150 verified sources for hearing about every new Codex feature, best practice, workflow and breakout project, early. Organized by band, tiered by signal strength, with the under-the-radar gems marked. Every entry was checked live for activity within the last six months.
BUILT 06-09-2026 · 11 RESEARCH SWEEPS · LIVE WEB CHECKS · LAST SYNCED 24-09-2026 12:04 EEST
What the daily sweep of the core sources turned up, newest first. One line per item; anything older than seven days drops off.
If you adopt nothing else, these twelve, in this order, cover most of the signal. Everything else on the page deepens a band these open.
OpenAI's own weekly digest of Codex and ChatGPT feature changes, with usage examples. The curated layer above the exhaustive changelog, and the single highest-signal page for Codex itself.
Watch → Custom → Releases on the repo. Every CLI version lands here, prereleases included, usually before any doc catches up.
The purpose-built official developer feed. Carries launches, limit changes and Codex announcements that never get a blog post.
Where new workflows, configs and regressions surface within hours. Pair the subreddit with community.openai.com/c/codex/37, where OpenAI staff answer.
The only place API behaviour changes and model availability are announced with dates. Everything you build on Codex sits downstream of it.
Auto-generated from the API spec, so a new model id, endpoint or beta header appears here at merge time, days before the release notes. The JS twin is often first to name a model in writing.
The most rigorous independent voice on LLMs and agents, publishing several times a day. Day-one significance analysis of a Codex change usually starts here.
The pre-release layer: feature-gate and app-string archaeology that reports ChatGPT and Codex features weeks before the changelog does.
New agent capabilities translated into reusable engineering patterns rather than demoed. Pair with AI Code King for a same-day hands-on verdict on anything that shipped.
The convention every harness now reads, kept in the open. Watch it alongside the community catalogs of Codex skills, subagents and plugins for vetted new tooling.
Where the engineering culture argues about Codex first, and where comment sections surface tools days before newsletters. Track by date via Algolia; scope the query, since the word names unrelated projects too.
The highest-signal long-form interviews in the space. New OpenAI product directions often get their first deep technical airing here.
The first-party surfaces. Four separate release-notes tracks exist: the Codex CLI on GitHub, the combined ChatGPT and Codex docs changelog, the API changelog, and the consumer help centre. Most people watch one of them.
The authoritative record of every Codex CLI change, usually live before any doc catches up: 0.153.0 on Sep 3, 2026 shipped Vim undo/redo and a plugin marketplace CLI, then hotfixes 0.153.1 through 0.153.4 landed within about a day. New models reach the release notes first, as with GPT-6 Astra appearing in the bundled model picker.
The combined product changelog covering the Codex CLI plus ChatGPT desktop and mobile apps; it logged CLI 0.152.0 through 0.153.4 across four September days alongside ChatGPT for iOS builds. Buried by a move: developers.openai.com/codex now redirects to learn.chatgpt.com, so older guides point at the dead path.
Every API change the Codex ecosystem builds on: the Sep 3, 2026 GPT-6 Astra entry includes migration limits such as no custom temperature and tool calling requiring the Responses API, plus async tool calling and mid-turn WebSocket steering shipped the same day. Codex model releases like gpt-5.2-codex and gpt-5.3-codex are announced here. platform.openai.com/docs/changelog redirects to this page.
The consumer-app release notes for the ChatGPT surfaces that host Codex; entries cover feature rollouts to Plus, Business, and Pro plans and retirements such as o3 leaving ChatGPT on Aug 26, 2026. ChatGPT-side agent changes show up here before they reach the developer docs.
The shutdown calendar that breaks Codex setups silently: the entire gpt-5-codex and gpt-5.1 and gpt-5.2-codex line was switched off on Jul 23, 2026, and reusable prompts, the Evals platform, and Agent Builder all end on Nov 30, 2026. The stated policy allows preview models to retire with as little as two weeks of notice, so a monthly check is self-defense.
The canonical launch surface, where GPT-6 Astra arrived on Sep 3, 2026 together with its system card and safety overview. Product, research, and safety posts land here together and nowhere else; subscribe via the feed at openai.com/news/rss.xml because the site itself is heavy to browse.
Staff posts that often beat the blogs: the Codex-focused OpenAI Build Week (Aug 26, 2026), DevDay going global, the WebMCP Challenge, and price cuts spanning API, Codex credits, and ChatGPT Work all ran here. Also where developer office hours get scheduled.
OpenAI's developer account, with the bio 'Official updates for developers building with Codex and the OpenAI Platform'. It announced Codex Security, posted a dated tease for Codex shortcuts ('July 15th'), shares weekly Codex quality-of-life threads, and broke the DevDay global tour news. Frequently the first public word on a Codex feature.
One merged feed of ChatGPT, API Platform, and Codex release notes on openai.com, each item linking back to its source doc, with its own feed at openai.com/products/release-notes/rss.xml. In September 2026 it carried the GPT-6 Astra rollout notes, new Codex browser policy controls, and healthcare plugins for Codex side by side.
Implementation deep dives that exist nowhere else: building the sandbox that enabled Codex on Windows (May 13, 2026), self-improving tax agents with Codex (May 27, 2026), and WebSocket steering in the Responses API (Apr 22, 2026). Deeply buried: the old openai.com/index/engineering URL is a 404, there is no nav entry, and its own feed at openai.com/news/engineering/rss.xml is barely watched.
Codex has its own four-component section on this page, and the incident feed at status.openai.com/history.rss logged Codex-specific trouble on Aug 11, Aug 20, and Sep 3, 2026 out of roughly 50 incidents since June. When Codex misbehaves, this page answers before support does.
Runnable recipes that show where OpenAI thinks agent work is going: 'Build iterative repair loops with Codex', the Codex prompting guide, workspace agent patterns, and a migration guide from the Claude Agent SDK to the OpenAI Agents SDK. Note cookbook.openai.com now redirects here.
The TypeScript sibling of the Agents SDK and the faster-moving track: v0.15.0 changed the default model to gpt-5.6-luna and added MCP v2 negotiation, and v0.14.0 brought Programmatic Tool Calling with AI SDK 7 support. One core maintainer signs most releases, so watch the pace.
The current docs home for Codex CLI, the IDE extension, Codex cloud, plugins, and enterprise admin; developers.openai.com/codex redirects here, worth knowing because most external guides still cite the old domain. It also hosts the What's new digest and the full changelog in one place.
A curated weekly digest of ChatGPT and Codex feature changes with usage examples, deliberately distinct from the exhaustive changelog: the Aug 17 to 21, 2026 edition covered the GitLab beta in Codex cloud, and Aug 10 to 14 covered importing from Claude Code and Cursor. The cheapest way to catch features that ship without an announcement post.
The official OpenAI server with roughly 855,000 members, whose own description names ChatGPT and Codex; OpenAI's help center documents it as the sanctioned community with a verification step. The fastest place to learn whether a Codex or API error is just you.
Free official courses including a dedicated Codex track and a workspace-agents course on directing agents and reviewing outputs. New course tracks here telegraph which surfaces OpenAI is pushing toward mainstream users, and the Codex course material tracks current product behavior.
Around 780 open roles, and the Codex postings telegraph unannounced product lines: current listings include Codex Enterprise, Codex Core Agents, Codex Cyber, and a Product Manager for Codex Security Controls and Partnerships. New team names here tend to show up in the product months later.
The most distinctive band on this radar: sources that see features in SDK releases, npm builds, the OpenAPI spec and app strings before OpenAI announces them.
New model ids and endpoints land in these release notes at merge time, before any blog post: v3.8.0 on 2026-09-03 added gpt-6-astra, and v3.3.0 quietly added named data-residency endpoints. Following the feed is the cheapest way to watch the API surface change as it happens; ten releases shipped between Aug 12 and Sep 3, 2026.
The JS twin of the Python feed and often the first place a model name appears in writing: v7.4.0 added the gpt-5.5 model id and tool metadata fields, v7.5.0 added Daybreak and gpt-5.6-cyber identifiers. If you build on the API or Agents SDK in TypeScript, this is where breaking changes and new params reach you days before the docs changelog.
OpenAI's machine-readable OpenAPI spec, auto-synced from upstream source, so new endpoints, params and schemas appear here before the docs mention them; commits on 2026-09-05 and 09-06 adjusted live-call event shapes and input item limits. At about 2.5k stars it is watched far less than the signal justifies, and the YAML diffs are scriptable for alerts.
Verbatim captures of Codex system prompts kept per model and per mode: GPT-5.6 Terra/Luna and Sol variants, Codex Spark, plan mode, auto-review and computer-use prompts, updated as recently as 2026-09-05. Watching the OpenAI folder shows how Codex instructions change between models and modes, which OpenAI never documents itself.
The app-string archaeologist: he digs new strings out of ChatGPT web, Android and desktop builds and posts the finds himself, in 2026 including a Pro Lite plan tier spotted in checkout code and ads strings that tech press covered days later. His weekly OpenAI and Anthropic recaps put each find in context, and he also posts on Threads and Medium under the same handle.
Aggregates in-app experiments, code-string finds and closed-test leaks across ChatGPT and its rivals, with leaks clearly separated from confirmed launches. It carried GPT-6 Astra output samples and alpha-access details in late August 2026 ahead of the Sep 4 launch, and covered new Codex context memory at rollout. Works as the followable inbox for what otherwise lives in scattered X posts.
An open-source database of model metadata keyed to what providers actually serve, run as the sst/models.dev repo. Its OpenAI rows already list GPT-5.3 Codex, Codex Spark and Codex Max variants with 2026-09-04 dated entries, so a new or quietly deployed model id tends to show up here fast, with context limits, pricing and capability flags attached.
The Agents SDK release notes surface agent-platform features as they ship: MCP Python SDK v1 and v2 support in v0.20.0, the ProgrammaticToolCallingTool in v0.19.0, and a default-model bump to gpt-5.6-luna. Default model choices here are a leading indicator of which models OpenAI treats as production-ready for agent workloads.
The extension publishes under internal build numbers that jump several times a week: 26.5901.22334 landed 2026-09-03 right after a run of 26.5825 and 26.5820 builds, with more than 11 million downloads on Open VSX alone. Build-number activity is often the first visible sign of a paired ChatGPT app release, and the extension offers a prerelease channel ahead of stable.
The registry entry where every build actually lands, platform packages included, with the alpha dist-tag visible in the registry JSON: 0.154.0-alpha.3 and its per-platform builds appeared 2026-09-04 alongside stable 0.153.4. Because the registry exposes publish timestamps programmatically, you can alert on version bumps within minutes of upload.
The openai organization's repository list sorted by creation date. New public repos appear there unannounced, and OpenAI's agent tooling has a habit of landing as fresh small repos rather than inside existing ones; a weekly glance catches incubated tools before anyone writes about them. There is no atom feed for org repo creation, so this is a manual check.
Tiered by how much they will change your practice: workflow-thinking channels first, then fast tutorials, then news wires and specialist angles.
The source itself: 2.04M subscribers, 763 videos. Launch films for Codex and GPT models, DevDay keynotes, and developer use-case shorts like Introducing GPT-6 Astra for developers (4 Sep 2026) land here before anyone has hands on them. If a new Codex capability exists, this channel demonstrates it first.
A working product engineer (T3 Chat, ping.gg) who pressure-tests every new agent and model within hours, about 493K subscribers. In the first week of September 2026 alone he covered the OpenAI Cursor ban story, a twelve-cent hundreds-of-PRs workflow, and Astra on day one. The fastest honest read on whether an OpenAI release actually changes how you build.
132K subscribers, just under a thousand videos, and a hands-on test of basically every coding-agent and model release the same day. His GPT-6 Astra benchmarks deep-dive asked whether it is still a good coding model the day after launch. The fastest way to decide if an OpenAI model or agent update matters for your stack.
About 228K subscribers and daily uploads; every major model drop gets a fully tested review the same day. His GPT-6 Astra IS AGI video, with use cases and tricks, landed the day of release. Consistent same-day OpenAI coverage when a GPT or Codex model ships.
Senior agentic engineering, about 147K subscribers, weekly. Runs ongoing series on software factories, agent sandboxes, and where to focus agents at the operating level; treats agents as an engineering discipline with structure rather than a tips channel. Under-followed relative to the depth on offer.
An AI engineer with 10-plus years in production, about 290K subscribers, long-form builds roughly twice a month. Real architecture, not demos: how I built one backend to run my company, document-review apps, full-stack GenAI stacks. Where you learn how agent systems survive contact with production.
Coding-agent practice rather than hype, about 224K subscribers and several uploads a week: 11 tiny coding agent fixes with a stupid amount of payoff, hooks that stop agents ignoring your rules, and an interview with BMAD's founder. Aimed at the actual hard part of Codex-style work, making the agent obey your conventions.
A Google Developer Expert in ML, roughly 200K subscribers, who was early with a full build-an-agent walkthrough for the OpenAI Agents SDK and keeps a 21-video Agent Frameworks playlist spanning OpenAI, Google ADK and LangGraph. Calm, code-first tutorials; also covers sandboxes for agent security. The video to watch when a new SDK version lands.
About 111K subscribers and 1,400 videos, among the fastest to publish a working walkthrough after any OpenAI release: GPT 6 Astra in 11 mins, and OpenAI Codex Just Replaced Humans Browsing. Short, business-friendly Python; good for quick verification that a new endpoint or Codex feature actually works.
Plain-language AI tool guides with about 337K subscribers. The Ultimate Guide To ChatGPT Codex for Everyday People (about 55K views, 2026) is the video to hand a teammate who is not a developer. Fast coverage whenever OpenAI ships a Codex or ChatGPT feature aimed at normal users.
About 1M subscribers and 783 videos; the Friday AI news recap with commentary and demos is the ritual, and he runs the FutureTools directory. His GPT-6 Astra Is Finally Here review came the day of release. A reliable weekly digest so an OpenAI announcement week never passes unnoticed.
About 633K subscribers; tests frontier models hands-on rather than reading headlines. He had early access to Astra and posted three separate test videos on launch day (What GPT-6 Astra Can Do, 3 Things to Try). Good signal on whether a new OpenAI model is worth switching your agent to.
The biggest dev channel on this list at about 4.2M subscribers. Pattern: a sharp ten-minute explainer within 48 hours of a release, like Did OpenAI actually build AGI? GPT-6 Astra first look (4 Sep 2026). Use it to catch and contextualize what you missed, not to learn workflows.
Fewer than 3K subscribers on a restarted channel, yet his Learn 100 hours of Codex in 18 minutes pulled tens of thousands of views in May 2026. A practitioner condensing hard-won Codex and agent-tooling habits into short videos, including how to properly set up agent tools. Sporadic but unusually dense.
82K subscribers and 646 videos of compact, no-fluff agent fundamentals: Hermes Agent desktop and custom skills, local model serving, weekly AI update roundups. Covers agent frameworks in ten-minute walkthroughs that bigger channels skip; quietly one of the most complete agent tutorial catalogs.
About 445K subscribers; the author of Simple Bench, a benchmark cited by The Economist and MIT Technology Review, so his capability analysis rests on his own measurements, like GPT 6 Astra, so good even OpenAI are worried (4 Sep 2026). A few videos a month; the context layer over release news.
About 232K subscribers, roughly weekly, doing research-grade breakdowns of what is under the hood: Kimi K3 architecture, RL training claims, and GPT-5.6 Sol benchmark-mining questions. For practitioners who want to reason about why a model behaves the way it does inside their agent.
About 326K subscribers and near-daily uploads; the fastest commentary wire on AGI-adjacent news, OpenAI included, like OpenAI just crossed a THRESHOLD (5 Sep 2026). Broad and high-volume rather than deep; useful as the alarm bell that something big dropped.
The English staples, then the genuinely underused edge: some of the deepest Codex write-ups are published in Japanese, Chinese and Korean, and code blocks and configs need no translation.
OpenAI's own forum category for the Codex agent, IDE integration and coding tools, with threads active within the hour of reading. The Codex rate limits discussion thread alone carries 632 replies and 38.8k views, and policy shifts like the 5-hour usage limit and the Codex for OSS program get argued here before anywhere else. An RSS feed exists at community.openai.com/c/codex/37.rss.
The community surface of the 121,950-star Codex repo, with 29 pages of threads across Announcements, Q&A, Ideas and Show and tell. OpenAI staff pin calls to action here, such as the help test the experimental Windows sandbox announcement, and ecosystem tools like CodexFuse 1.2.0 and the Click plugin debuted in Show and tell in August 2026.
The subreddit devoted specifically to Codex CLI, the IDE extension and Codex cloud, around 168,000 members by the latest tracker count. Reactions to quota and pricing changes land here within hours, and threads on running Codex against an existing ChatGPT quota surface community workarounds early.
15,586 open issues, and the five most recently updated were all touched within the final hour before reading, covering GPT-5.6 aborting tools.exec_command and codex-code-mode-host 0.147.0 SIGTRAP crashes on WSL2. The fastest read on what broke in the newest CLI or IDE extension release.
The general AI coding subreddit, roughly 390,000 members per GummySearch's count, where Codex changes get debated directly against every rival tool. Codex staff have hosted AMAs here, making it one of the few places to put questions to the team outside the official forum.
A saved by-date search over all of Hacker News via Algolia, scriptable through its API, which is the practical way to watch HN for Codex without camping on the front page. Fresh catches in the first week of September 2026 included Show HN: Layerz and the story on OpenAI agents abusing ntfy.sh as a pub/sub channel.
Practitioner blog posts tagged codex, with items published the day of reading, among them Why ChatGPT/Codex Desktop Had No Window After an Update on Windows and Tracking Codex Astra costs. Deep workarounds like repackaging the official Windows Codex MSIX into a Linux .deb appear here first.
The small invite-based aggregator returns a thin but steady stream of coding-agent stories matching codex, with the newest from the last week of August 2026, including JetBrains bringing IntelliJ intelligence to agentic coding flows and a provider-agnostic agent toolkit. The draw is the comment threads, where working engineers dissect coding-agent behavior with less noise than HN.
Japan's heaviest developer blog platform by volume, with 770 Codex articles, 3 books and 34 scraps, and an RSS feed that updated the day of reading. Japanese engineers document CLI internals, config tricks and cost experiments here that English feeds and search never index.
147 tagged posts with 43 followers and a created-date feed that showed new items the day of reading. Its trend list holds heavily liked Japanese deep dives, like Spec Driven Codex at 95 likes and the post driving Codex from Claude Code over MCP at 90 likes, work that does not surface in English search.
China's largest developer blog hub, whose Codex tag page advertises fresh Codex content daily and carried a run of 2026 guides and comparisons. These cover corners English posts skip, such as the CLI dropping Chat Completions support for the responses wire API in February 2026 and wiring Codex to third-party gateways on macOS and Windows.
Engineers who publish repeatedly about agentic coding practice, where the vocabulary of the field gets coined months before it is mainstream.
Tracks every Codex release and model change within a day on an active codex tag, including the September 2026 post explaining why the ChatGPT desktop app bundles LibreOffice and Poppler runtimes for Codex. He reverse engineered the Codex CLI auth to run its models from other tools and shipped the llm-openai-via-codex plugin. His February 2026 essay How I think about Codex is still the clearest mental model of where Codex fits.
Turns emerging agent practice into named, usable frameworks. Practical Loop Engineering of August 14, 2026 covers Codex and Claude loop primitives including the Codex goal command, Agentic Autonomy Levels compares Codex Goal mode against peers, and his GPT-5.6 Sol notes cover running Sol Ultra inside Codex.
Coined the Ralph loop, the named pattern for running a coding agent in a continuous cycle of reading, poking, and committing, and generalized it in Everything is a Ralph loop. Active through July 2026 with practitioner pieces such as engineer away the slop. Also gives away a free how to build a coding agent workshop that covers Codex alongside its peers.
Builds and documents Superpowers, the skills framework that now spans Claude Code and Codex, with Codex support laid out in the Superpowers 6 post of June 15, 2026. Publishes results from his own builds, such as vibe coding a C compiler able to build SQLite in August 2026. His blog tagline names agents, coding, Claude, and Codex as core beats.
High volume hands on testing of Codex CLI changes, for example Codex untrusted Approval Policy Retired, What to Change, covering Codex CLI 0.149 on August 25, 2026. Runs blind Claude versus Codex comparisons on the same tasks rather than repeating vendor claims.
How I use AI in 2026, from August 17, 2026, reports round robin builds with completely vanilla Codex and Claude Code, a ChatGPT Pro plan never maxed out, and about 500 dollars a month of spend, all from his own logs. Followed it with The Harness Is the Company on August 24, 2026. Over 3,000 subscribers read these usage grounded reports.
Short, concrete reports from real sessions, like A week of using Codex more than Claude from August 21, 2026, which observes that Codex sticks to the task and halts while Claude guesses what you might want. Where Six AI Coding CLIs Store Your Session Logs of August 6, 2026 is the reference on Codex session log locations. Clearly under followed relative to the output.
Auto-research with codex, July 8, 2026, documents running GPT-5.5 through Codex goal loops over 14 days, with more than 1,500 submissions and a 232x kernel speedup that ranked 12th of 183 on the GPU Mode qr_v2 leaderboard. One of the clearest public records of what Codex goal mode autonomy actually produces.
Codex-maxxing from May 10, 2026 treats the Codex app as a knowledge work operating system, with durable megathreads, a 30 minute Chief of Staff heartbeat, Obsidian as memory reviewed through diffs, and Goals such as porting Python Rich to Rust while passing the original test suite. A long time OpenAI ecosystem educator writing from his own workflow.
Builds open source tooling around parallel coding agents and publishes the process. Swarm, April 2026, is a Rust and GTK4 Linux app that runs Codex and other CLIs in parallel over git worktrees with one click pull requests, and Chimera, July 2026, sandboxes agents in userspace.
GPT-5-Codex is a better AI researcher than me documents handing a GPU research loop to Codex with the danger full access sandbox, restarting the agent roughly every one million tokens. Keeps a steady 2026 stream of essays on working with models, such as How to protect yourself from workslop on September 2, 2026, all grounded in his own experiments.
Repos to watch, the frameworks that define practice, and the trackers that catch breakout projects on day two of their existence.
The source repo for the Codex CLI, its IDE extension, and the app protocol. About 122,000 stars and 18,700 forks, with release rust-v0.153.4 shipped 2026-09-04 and commits the same day. Watching releases and docs changes here shows new model support, sandbox, and config features days before any blog post.
OpenAI's official recipe repo, about 75,800 stars, pushed 2026-09-05. New recipes track each API surface, including Responses API and Agents SDK patterns, so the commit feed works as a quiet changelog for the platform behind Codex.
The Agents SDK itself, about 29,200 stars, pushed 2026-09-05. Model updates, tracing changes, and new tool primitives for agent development arrive as releases here; reading the release notes tells you what the SDK can do before tutorials catch up.
The AGENTS.md spec and adopters list, about 24,200 stars, pushed 2026-08-25. The site counts over 60,000 open source projects using the format and lists Codex, Cursor, Gemini CLI, Copilot's coding agent, and Zed as compatible; commits here are how the instruction file format Codex reads actually changes.
A live directory of 29,738 repositories tagged codex, sortable by recently updated, with language filters such as 9,325 Python and 4,742 TypeScript repos. The fastest way to size the ecosystem and spot new entrants; right now it surfaces cc-switch at about 131,000 stars and a UI/UX skill repo at about 125,000.
The daily and weekly breakout lists for all of GitHub. Codex-adjacent tooling does land here: the 2026-09-06 daily list carried mattpocock/skills, a collection that installs into Codex, at about 2,200 stars gained in one day. Checking it daily is how you catch such repos during their first week of lift.
OpenAI's security scanning CLI and TypeScript SDK built around Codex, about 10,500 stars, created 2026-07-13, pushed 2026-09-06. The newest official piece of the Codex family; its README already covers bulk Docker scanning setups and alternative inference providers, so watching it shows where OpenAI is taking agent-driven security review.
The official Python SDK, about 31,600 stars, pushed 2026-09-05. Codex and Agents SDK work is Python-first, so new model identifiers and API parameter types appear in this client before almost anywhere else; its release notes are the de facto API changelog.
The official JavaScript and TypeScript SDK, about 11,200 stars, pushed 2026-09-06. Serves the same changelog function for the TypeScript side of the ecosystem, which is where the Codex CLI's npm package and the codex-security SDK live.
The TypeScript sibling of the Agents SDK, about 3,800 stars versus about 29,200 for the Python one, pushed 2026-09-05. It ships the same agent, tool, and tracing primitives for web and voice agents, and is far less followed than its value justifies.
A skills collection with about 254,000 stars, pushed 2026-09-04, that installs into Codex via npx skills@latest add mattpocock/skills. Its README documents failure modes seen across Claude Code, Codex, and other coding agents, and new skills appear here before they get copied into Codex-specific collections.
A desktop switcher for Codex, Claude Code, Gemini CLI, and other tools, about 131,000 stars, pushed 2026-09-06 with a v3.16.1 release note. It manages multiple Codex provider accounts and config files, so its changelog doubles as a map of what config surface Codex exposes.
A usage and cost reporter for local agent logs; the README documents npx ccusage codex daily alongside Claude Code, OpenCode, and others. About 18,400 stars, pushed 2026-09-06. When OpenAI changes logging or rate limits, fixes land here fast because it parses those session files directly.
A macOS menu bar app showing usage stats for Codex and Claude Code without login, about 21,000 stars, pushed 2026-09-06. It reads local session logs and keeps history across limit resets, which makes it the practical early warning for weekly limit and pricing changes.
A kanban board that runs Codex, Claude Code, and Gemini CLI agents in parallel, about 28,000 stars. Honest caveat: the last push was 2026-04-24, inside the six month window but quiet, so watch it for revival or fork momentum rather than steady updates.
A curated directory of practical Codex skills across the CLI and API, about 16,300 stars, pushed 2026-07-26. The best one-stop index of the skills layer of Codex; new entries here usually point at new repos elsewhere, which makes it a discovery layer more than a static list.
A collection of over 130 specialized Codex subagents, about 6,100 stars, pushed 2026-09-04. Copy-paste agent definitions for review, testing, and documentation work; watching additions shows which delegation patterns the community has found to actually hold up.
A created-after date filter that catches repos too new to have meaningful stars. The codex query with created greater than 2026-06-01 returned about 77,600 results, including openai/codex-security at about 10,500 stars within two months of creation. Bump the date forward monthly and rerun to keep it a fresh feed.
A curated list of Codex and ChatGPT plugins, skills, and resources, about 940 stars but pushed 2026-09-06. Small following next to the bigger lists, yet it tracks the plugin and marketplace corner of the ecosystem that the skills-focused lists skip.
A design intelligence skill for coding agents, about 125,000 stars, pushed 2026-09-06, installed with uipro init --ai codex and listed as supporting Codex CLI in its skill mode list. One of the most starred repos tagged codex, and a case study in skills becoming the distribution format for agent capabilities; fits you if you ship UI through Codex.
Tool protocols and file conventions flow downstream into every Codex workflow. The upstream signal, spec proposals and maintainer channels, is where a shift is visible months early.
The MCP spec repository is where every protocol change Codex depends on lands first, as a pull request against the TypeScript schema that the project declares the source of truth. The tree carries the 2026-07-28 snapshot alongside the draft, with about 75 pull requests open at once in September 2026. Watching the PR stream shows breakage, like the roots, sampling, and logging deprecation, long before a dated release reaches your tools.
The site for the convention Codex reads natively at the start of every task: plain Markdown, no schema, nested files where the closest one to the edited file wins, and explicit user prompts override the file. It counts over 60k open-source projects carrying one, with supporters from Codex and Gemini CLI to Cursor, Copilot, Junie, Zed, and Devin. Now stewarded by the Agentic AI Foundation under the Linux Foundation, it is a shared contract rather than an OpenAI-owned format.
OpenAI's universal plugin system for ChatGPT and Codex: a plugin bundles skills, MCP servers, and optional UI, and publishes once to a shared directory governed by public MCP server review requirements. The docs include a guide for converting Claude Code plugins into OpenAI-submittable ones, which states the interop frontline outright. Monetization via a checkout API and flows like ChatGPT's Reserve show where plugin-provided tools are being pushed.
The numbered pipeline for major MCP changes, with statuses from Draft through Final; 41 SEPs are final. SEP-2577 deprecated roots, sampling, and logging, and SEP-2663 stabilized the Tasks extension for long-running agent work. Items sitting in Draft or In-Review are the months-early view of what the next snapshot will force through every MCP client, Codex included.
The argument floor beneath the spec: September 2026 threads include a tool pagination proposal, a complaint that the pagination docs are misleading, and working group meeting notices. Bigger fights this year covered agent identity and delegation for tool calls, with 36 comments, and a Tool Bill of Materials for supply-chain provenance. Positions staked out here become the SEPs your servers must later conform to.
A subscribable iCalendar feed of every MCP working and interest group: Transports, Agents, Triggers and Events, MCP Apps, Registry, Fine-Grained Auth, and Skills Over MCP meet weekly or biweekly. This is where protocol direction is visible earliest; the Agents WG, chartered 2026-08-04 and led from AWS and Microsoft, is pushing Tasks from extension into the core protocol. Each group also runs a Discord channel where the drafting actually happens.
The official directory for discovering MCP servers, in preview since 2025-09-08 with its API frozen at v0.1, and its repo keeps publishing continuously. The Registry WG meets biweekly in public on the MCP calendar. Track it to see whether servers you ship or depend on are discoverable by conforming clients once GA lands.
The spec repository behind the convention, with about 80 open pull requests adding agents and debating compatibility. A 2026-08-25 commit updated the Codex logo, and March 2026 commits added Junie and Augment Code and refreshed the Gemini CLI instructions. Any change to how agent instructions are discovered or scoped would start in these issues.
The definitive map of Codex as an MCP client: codex mcp add, project-scoped config.toml, the /mcp command, Streamable HTTP with OAuth including CIMD and DCR, per-tool allow lists, output_token_limit, and an experimental remote environment for stdio servers. It documents plugin-provided MCP servers under plugins.
How ChatGPT, deep research, and the Responses API consume remote MCP servers: the required read-only search and fetch tool contract, citation behavior that depends on a non-empty url in results, and the mcp tool type with server_label, allowed_tools, and require_approval. Auth guidance centers on OAuth with Client ID Metadata Documents and tracks the 2025-11-25 MCP auth spec. Custom servers attach at chatgpt.com/plugins with developer mode enabled, so this page defines the contract your server must pass to reach OpenAI's surfaces.
Zed's open protocol for editor to agent communication, JSON-RPC over stdio that deliberately reuses MCP's wire types. Codex CLI connects through an adapter maintained by the ACP project, alongside Gemini CLI, GitHub Copilot in preview, Cursor, Claude via Zed's adapter, and dozens more. ACP v2 is in draft and the Rust and TypeScript SDKs reached 1.0, so if agent-editor interop consolidates, this is the second standard to watch beside MCP.
A server and client directory with an editorial layer, run by MCP steering committee member Tadas Antanavicius. Recent essays propose agentic MCP configuration as an answer to tool overload and walk through pricing a paid MCP server. Server submissions are paused during an ingestion revamp, so read the index knowing it is partially stale.
Where OpenAI leadership and the Codex team go long-form, and the structured paths for going deep on a feature the month it matters.
The developer conference where Codex, AgentKit and platform releases have shipped: the 2026 edition runs September 29 at Fort Mason in San Francisco with a Sam Altman keynote at 10 a.m. The keynote is livestreamed free and breakout sessions are recorded and posted on openai.com afterward, so remote practitioners get the sessions, not just the headlines. Caveats: in-person applications are already closed and the cadence is annual, with DevDay Exchanges in cities like Bengaluru, Tokyo, Seoul and London later in 2026.
OpenAI's own long-form show, hosted by Andrew Mayne with interviews of the people inside the company, including episodes like Inside the Model Spec and conversations with researchers such as Alexander Wei and Lijie Chen. It is 23 episodes in, updated semiweekly, and still barely known relative to its access, which is why it is a gem. For a Codex practitioner this is the closest thing to hearing the Codex and platform teams think out loud between releases.
The AI engineer show that gets the OpenAI platform team on the record: the July 28, 2026 episode interviewed Akshay Nathan on Codex going from 0 to 10M users and the merger of Codex into ChatGPT Work, including harness details like shared agent scaffolding and Memory V3. It also runs emergency live sessions on DevDay days, with an exclusive recap interview with the team that shipped AgentKit. This is where practitioner reaction and insider interviews meet.
OpenAI's official course platform, free per its help center, with a self-paced pathway of AI Foundations, Applied AI Foundations, and Agents and Workflows, plus OpenAI course completion certificates. Deployment material updated June 2026 shows it is maintained, not a launch relic. Caveat: it targets end users and knowledge workers more than engineers, and there is no Codex-specific course, so treat it as the source for how OpenAI teaches agent concepts, not developer deep dives.
The catalog to search when a capability matters enough to learn properly, shipping several free short courses a month on agents, code agents, and workflows. It carries the courses built with OpenAI, notably ChatGPT Prompt Engineering for Developers taught by OpenAI's Isa Fulford with Andrew Ng, still live and free. Caveat: the OpenAI co-created courses date from 2023 and 2024, so newer agent courses are partner-built and vendor-neutral.
The main practitioner conference for agent tooling, held June 29 to July 2, 2026 at Moscone West, with every talk professionally recorded and posted free on the AI Engineer YouTube channel. The 2026 page links straight to the recordings, and past editions carried the tracks practitioners care about, like the MCP track and agent harness sessions. It is where the ecosystem around Codex, MCP and agent SDKs shows its working, not just its demos.
Kevin Roose and Casey Newton get OpenAI leadership in the room early, from the Hard Fork Live sit-down with Sam Altman and Brad Lightcap to prompt coverage of every model event; a July 24, 2026 episode dug into OpenAI models going rogue. Weekly and reliably first with a mainstream interview after any OpenAI news wave. It is strategy and narrative rather than API detail, so pair it with a practitioner source.
Alex Volkov's weekly live roundup, presented by Weights and Biases, is the fastest structured take on every OpenAI release, and it gets OpenAI staff on directly: the July 23, 2026 special featured OpenAI's Romain Huet on Codex crossing 5M users and GPT-5.6. It covered GPT-5.6 landing mid-show on July 9 and ran a GPT-6 Astra deep dive on September 3, 2026. A community show with modest reach and unusual access, which makes it a genuine gem.
Elad Gil and Sarah Guo interview the operators building on top of frontier models and have had OpenAI leadership on, including CPO Kevin Weil on product strategy and the roadmap around ChatGPT and the API. The show is active weekly in 2026, though OpenAI guests are occasional among chip and investing themes. Useful for hearing how serious builders read each OpenAI move, with the caveat that you wait for the episodes that matter to you.
YC partners talking to founders about how work actually changes, and recently that means coding agents: the February 6, 2026 episode featured Calvin French-Owen, a former engineer on OpenAI's Codex team, comparing Codex and Claude Code and explaining context management in practice. Companion pieces like Tokenmaxxing cover how top builders run agent workflows across hundreds of engineers. Most episodes are general startup talk, so subscribe for the agent-focused runs.
Multi-hour interviews with the people deciding AI's direction, including repeated Sam Altman sit-downs, and freshly topical agent material: the September 1, 2026 episode with Ajeya Cotra went inside the OpenAI agent swarm that hacked Hugging Face, days after his own explainer on the incident. The New York Times profiled it in April 2026 as the show where you eavesdrop on the AI elite. Caveat: OpenAI-specific episodes are occasional among compute and science interviews.
Where model and harness changes register before marketing does, including stealth checkpoints under codenames.
The benchmark built for terminal agents: Stanford, Harbor, and the Laude Institute run models in real shells and publish resolution rate, dollar cost, and tokens side by side. A Codex harness entry dated Sep 3, 2026 sits near the top, so it shows what the Codex stack costs per solved task, not just accuracy. Good for agent and harness comparisons; near-tied frontier scores are its saturation point, which is why harder task sets keep shipping.
The reference agentic software engineering benchmark every frontier launch cites, testing real GitHub issue fixing. The February 2026 re-run refreshed scores against current models and added third-party scaffolds including Codex and Claude Code, so you can compare models under one scaffold instead of each vendor's own. Top Verified scores now cluster in the mid-90s, so it barely separates frontier models and shines mainly on new or mid-tier entries.
Crowdsourced head-to-head voting where providers, OpenAI included, have historically parked unreleased models under codenames before launch; gpt2-chatbot surfaced here ahead of GPT-4o. Now rebranded from LMArena with an Agent tab alongside the text and coding boards, so model-watch happens on both. Good for catching stealth launches early; it measures chat preference, not terminal agent skill, and codename sightings need community corroboration before you trust the guess.
Ranks models by actual tokens routed through OpenRouter, so coding traffic shifting to a new GPT or Codex release shows up within days, not at the next benchmark cycle. Data was current through September 5, 2026 on visit, published under CC BY 4.0. The page itself says it measures adoption, not quality, and it only sees OpenRouter traffic, so first-party API volume is invisible here.
A neutral third party that re-runs Terminal-Bench v2.1 (89 curated tasks) under one fixed harness, Terminus 2 in an e2b sandbox with pass@1 over three repeats, removing the scaffold noise that plagues vendor-reported scores. GPT-6 Astra sits at 89.9 percent against Claude and others on the same setup. Good for apples-to-apples agent scores on day one of a release; with the top already near 90 percent, frontier separation is shrinking.
A radar only works if the signal comes to you. Three mechanical moves, then a sustainable routine.
Every line below was fetched and returned a valid feed on 06-09-2026. The companion Feed Builder covers the sibling Claude radar.
status.openai.com/history.rss: incidents within minutes, Codex has its own componentsgithub.com/openai/codex/releases.atom: every Codex CLI releasegithub.com/openai/openai-python/releases.atom: API surface changes at merge timegithub.com/openai/openai-node/releases.atom: the JS twin, often first to name a modelgithub.com/openai/openai-agents-python/releases.atom: agent harness changesopenai.com/news/rss.xml: the newsroom, launches and system cardscommunity.openai.com/c/codex/37.rss: the forum's Codex categoryhnrss.org/newest?q=codex: every Hacker News Codex storysimonwillison.net/atom/everything/: the daily chroniclereddit.com/r/codex/.rss · zenn.dev/topics/codex/feed: the community firehose, English and JapaneseEvery source on this page is pollable. The whole radar can also run as a scheduled agent: a sweep of these feeds distilled into a digest delivered to you, so the checking itself costs zero attention. That is what the fresh-signal section at the top of this page already is.
Compiled 06-09-2026 from an eleven-band research sweep, run headless on the cheap rung of the model ladder and checked here. Every entry was verified live for activity within the last six months at compile time; cadences, star counts and follower counts are as of that date. Tier and gem labels are editorial judgments about signal per minute for a practitioner who wants new ideas early. Sources known to have caveats carry the caveat inline.
People and feeds: X, Bluesky, LinkedIn, Twitch, TikTok
The OpenAI insiders first, every handle verified, then the independent power users, then the networks most developer lists ignore entirely.
The engineering leader running Codex and ChatGPT, per WIRED. He posts user numbers and operational detail the blog never carries: Codex and ChatGPT Work hitting 8M active users, the Astra rollout to Pro and Business, and usage-limit resets while the team investigated. Most follow lists name Altman and Weil but miss the person who actually ships Codex.
Codex product lead at OpenAI (his bio says just 'Codex @OpenAI'). He shares power-user workflows from inside the team, like the power user's guide to Codex walkthrough, and gives the growth and roadmap framing in interviews such as Lenny's Podcast on what is coming in 2026. Earliest read on what the product team is optimizing for next.
Head of Developer Experience at OpenAI, explicitly working on Codex and the API. He demos new surfaces live before docs catch up and drops real numbers in interviews, telling ThursdAI that Codex reached 5 million users and previewing GPT-5.6. Follow to see the DX team's own example apps and API patterns as they land.
OpenAI's Chief Product Officer, who announced the Codex relaunch and fronts DevDay. He frames product bets in public (his posts on humans plus AI making new math discoveries show the direction he is pulling the platform). Launch-day and roadmap signal straight from the person who greenlights Codex features.
Announces model and product drops on X first, often ahead of or alongside the blog post, and replies directly to user complaints about Codex limits and pricing. His replies are where OpenAI's posture on developer feedback becomes visible days before any official change.
First-stop analysis of every Codex model: notes on GPT-5-Codex via the API, notes on GPT-5.1-Codex-Max, and his inventory of Codex's 232 tool interfaces. The Bluesky feed carries the links, first impressions, and reply threads with other practitioners a day or more before the polished blog post lands.
OpenAI president, now also leading product strategy per WIRED, and one of the faces of the original Codex launch video. His 2026 posts explain the vision ('Codex is for everyone, for any task done with a computer') and how OpenAI is retooling its own teams toward agentic development, context you will not get from release notes.
Chief Research Officer who publicly credited the GPT-5-Codex team by name (Andrey Mishchenko, Katy Shi, Hanson Wang and others), which is how you learn who actually builds Codex models. He was also an author on the original 2021 Codex paper. Sparse poster, but each post marks a capability shift.
The main account breaks the biggest Codex news first, such as the August 2026 open-source Codex Security CLI announcement. Needed as the anchor feed; the developer account supplements rather than replaces it.
The researcher behind OpenAI's reasoning models, the line of work GPT-5-Codex was tuned from. His April 2026 post arguing extra reasoning tokens are overrated, and his critique that benchmarks ignore the cost of thinking, shape how to read every Codex model release. Posts rarely but each one moves the discussion.
Ex-OpenAI founding member whose 2026 threads on AI writing most of his code, and on using models as thinking partners rather than code generators, define how practitioners interpret agent capability shifts. Not Codex-specific, but his takes frame what a coding agent should become, and he comments on OpenAI releases within hours.
Tracks the agentic coding race with actual numbers: his GPT-5.6 post flagged Codex possibly overtaking Claude Code as usage jumped about 1M users in a day to 6M plus, and he flagged the most important chart in the GPT-5-Codex release. His Latent Space interviews put OpenAI builders on record at length.
Runs cross-model coding agent evaluations and says what the scores mean, declaring that GPT-5.3-Codex killed the coding agent as a benchmark category. He shares his actual task prompts (picked up by Agentic Engineering Weekly), which is rarer and more useful than another leaderboard. An Amp/Sourcegraph engineer, so his Codex takes are competitor-vetted, not fan posts.
Got early access to GPT-5-Codex and posts same-day verdicts; his argument that folding the Codex app into a ChatGPT toggle misread developers drew a direct response from the OpenAI dev team, and community credit for getting a beloved feature restored. The X feed is the takes, his t3.gg videos are the follow-up deep dives.
Writes a long X article on essentially every Codex-adjacent drop: GPT-5.1-Codex-Max reactions, quoting the GPT-5.3-Codex system card on arbitrary code execution, noting GPT-5.4 landed in API and Codex, and reviewing GPT-5.6. He aggregates what everyone posted during a launch, so you miss less signal on the noisy days.
Alex Volkov's weekly live X Spaces plus newsletter, dedicated largely to OpenAI news. It is where Romain Huet said Codex hit 5 million users and teased GPT-5.6 before the announcement. The Spaces bring OpenAI staff and power users into one room every Thursday, so launch-week reactions get settled in public.
A Belgian technical coach publishing a weekly LinkedIn roundup of practitioner material on Codex and other coding agents, running steadily through July 2026; a recent edition surfaced Thorsten Ball's actual task prompt. Almost no developer follow list includes LinkedIn at all, which is exactly why this lands in your feed days before the X echo chamber repeats it.
swyx and team's show account, posting long interviews with the people who build the tools, including 'Codex from 0 to 10M Users: Building ChatGPT Work' with OpenAI's own engineering lead. The X feed also carries their AI News digest tracking Codex usage growth. Where you get an hour with a builder instead of a launch tweet.
A working engineer running Codex on odd real jobs, like converting DMX fixture files from hog4 to ma2 format, and posting what breaks. His feed is the unfiltered practitioner diary the official channels cannot have, plus deep dives on agent swarm architectures. He also teaches how to build your own coding agent on GitHub.
Former Netflix engineer who reacts to Codex model launches live, like his GPT 5.3 Codex stream the day it shipped. The value is the skeptical working programmer's first hour with a new model, including what he keeps and deletes, which no written review gives you. Clips of the reactions circulate on X within hours.
636K followers and 8.9M likes on short how-tos for AI agent tools, including 'Codex is The NEW Best AI Coding Tool' and Agents SDK walkthroughs. He also runs his own content business on Codex, per his Creator Economy interview. The fastest format for seeing a new Codex feature used start to finish in under a minute.
115.8K followers on TikTok with 1.3M likes, cross-posting the 100-seconds explainers from a 3M subscriber YouTube channel. Covers Codex launches in plain, fast form for developers who want the what-changed without a 40 minute stream. One of the few serious dev channels that treats TikTok as a real surface.
Wharton professor who runs Codex projects side by side with Deep Research and agent tasks and reports what actually survives contact with real work. His Opinionated Guide to Which AI to Use now carries explicit Codex recommendations, and his 2026 posts on agentic systems at work reach an audience the X dev circle never sees. The best LinkedIn-native follow for Codex practitioners.