Codex

@thsottiauxSep 25

Known affiliation

We are aware that codex is down and are working hard to bring back normal service.

@rauchgSep 25

We're helping world-class organizations like @klaviyo build agentic deployment platforms. ① Connect every agent (claude, codex, cursor…) ② Configure SSO via their IDP (Okta, Entra) ③ Now everyone can cook, securely® What happens to SaaS once every company sets this up? I get this question a lot 😁 First, the data must come from somewhere. These apps only work if they can be infused with business data. This is why we see the big enterprise SaaS players suddenly prioritize shipping CLIs and MCPs. Or dust off APIs they were neglecting for years. The new 'procurement bar' will be how ergonomic your product is for *agents*, rather than humans. How easily they can navigate your ontology and work with your data. Once that's in place, there's a long tail of SaaS applications that I suspect will never be bought again. They'll be generated. They'll be more secure, more performant, more modern, and tailored to each company's and employee's needs.

@mvanhornSep 25

Introducing Agent Tincan. Let your AI agents ask each other for help. Your agents each have a superpower. Grok Bot is always on in the cloud. Muse makes phone calls and has access to Meta. Instinct runs in iMessage. Codex and Claude Code have your code. ChatGPT and Claude have your chats. But they can't talk to each other, so you're the copy-paste. Now they can, all securely over Tailscale. 📞 Ask Grok Bot to have Muse call the restaurant 🖼 Get an image from ChatGPT, from any agent 🔎 Pull a detail out of an old Claude chat 🛠 Talk to Codex on your primary development machine Runs on your own private Tailscale network. No open ports, nothing leaves it. Open source. Setup is agent-first: paste one message into your always-on agent and it sets up the rest.

@iannuttallSep 25

I love @AmpCode and have been using it for 90% of my coding tasks in the last few weeks. But I really wanted to use my Mac Mini for computer use because it's already logged in to Chrome and has all my apps. Turns out you can just ask Amp to build that for you! Amp does have it's own computer use but for my use at least it was a bit slow, I couldn't paste passwords properly, and it wasn't easy to get it logged in to all my Chrome tabs. So I chatted with Puck, their assistant who is becoming more useful to me than Instinct and Grok Bot, and figured out a way to do it: 1. Set up an always-on Amp runner on the Mac Mini so an Amp thread can run there 2. A @herdrdev session with an open pane that lets Amp prompt it and reply 3. A personal Amp skill that tells agents in an Orb to farm out GUI and browser tasks to the Codex CLI on the Mini, via Herdr Amp can now run `codex --yolo` on my Mini with GPT-6 Luna High and chat back/forth to do whatever I need. I just used it to set up a new Google Oauth and Github Oauth app, as well as change some DNS records on Cloudflare. I have the ChatGPT Chrome extension installed on the Mini, Codex is the best by far at computer use and Luna is incredibly capable and cheap. Every task gets a fresh Codex instance so nothing carries over once Amp is done with it. I also have Jump Desktop so if there is a need to enter a password or 2FA code or a captcha I can use it on desktop or phone and then Amp tells Codex to continue. Here's the prompt and skill to give it your Amp agent: ``` Set up this Mac so my Amp agents in orbs can hand it browser and desktop tasks. I've added the delegating-computer-use skill (SKILL.⁠md and config.json). Read it first. Steps 1. Check Amp is installed and signed in, in the macOS account I use for the desktop. 2. Keep an Amp runner running with a user LaunchAgent that starts at login and restarts it if it stops. I use `amp --no-tui --runner-id <id> --remote-control-terminal --log-file <path>`. Check `amp --help` for the flags on this version. 3. Install Herdr (herdr.⁠⁠dev) and save its agent skill (`herdr --skill`) to a file. Create a dedicated Herdr session and workspace for the GUI worker with one shell pane. Don't touch my normal terminal panes. 4. Check Codex is installed, signed in and can use computer use. 5. Fill in config.json with the runner ID, working directory, Herdr session, workspace label, worker name, lock directory and Herdr skill path. Keep the model, reasoning effort and launch args unless I say otherwise. 6. Stop and tell me if something needs a password, a macOS permission prompt or a setting only I can change. 7. Give me a harmless read-only test I can send from an Amp orb thread, like opening example.⁠com in Chrome and reporting the page title.``` SKILL.⁠md ````markdown --- name: delegating-computer-use description: "Runs signed-in browser and native desktop tasks through a temporary high-mode Amp runner thread and a fresh Herdr Codex worker. Use when an Orb needs the user's local sessions or native GUI tools." --- # Delegate computer use Read bundled [config.json](config.json). Use the configured runner for local sessions and native GUI tasks. If unavailable, report the blocker. Never silently fall back to Orb Desktop or another runner/session. ## From an Orb 1. Call `list_runners` once; confirm `runner_id` is online and serves `working_directory`. 2. Create one temporary thread with `executor: "runner"`, configured `runner_id` and `working_directory`, `agent_mode: "high"`, and `intent: "environment-access"`. Omit `archive_when_done`. Tell it to load this skill, do the work itself, and create no further threads. 3. Include a stable request ID, this thread's reply URL, the task, authorization limits, and safe evidence needed. Require replies via `send_thread_message`; do not also use `wait_for_threads`. Save the returned runner thread URL. 4. Send all follow-ups, including completed human handoffs, to that same thread. Do not replace a blocked worker or replay an uncertain action. Already on the configured runner for this request? Execute the steps below directly. Only a temporary task thread may archive itself, never the originating project thread. ## On the runner Read `herdr_skill_path` and its worker reference. Use only the configured dedicated Herdr session/workspace (never an empty/default session) and live pane IDs. Only one GUI worker may run at a time. 1. Create the lock's parent if needed, then atomically claim `owner_lock_directory` with `mkdir`. On failure, report the conflict without touching the worker or lock. Never steal a live lock. Empty or old locks are not proof of abandonment; dead-owner recovery needs explicit approval and confirmation the owner cannot resume. 2. Write `owner.⁠md` inside the claimed lock: unique ownership token, request ID, owner/reply thread URLs, scope, and action state. Add pane/process identity after launch. 3. Discover and reuse the persistent workspace and shell pane. Create missing resources with `--no-focus` only if authorized. Require the shell in the foreground and no existing worker; otherwise retain the lock and report the conflict. 4. Start a fresh interactive Codex process/conversation for each new request. Use configured model, reasoning effort, and launch arguments; defaults are GPT-6 Luna high with mandatory `--yolo --no-alt-screen`. Verify actual launch state. Never substitute safer approval flags, use `exec`/`resume`/`--last`, or start a second worker after a timeout. `--yolo` does not expand task authorization. Use these commands with values from config and the discovered pane ID: ```sh herdr --session <herdr_session> workspace list herdr --session <herdr_session> pane list --workspace <workspace-id> herdr --session <herdr_session> pane process-info --pane <pane-id> herdr --session <herdr_session> agent list herdr --session <herdr_session> agent start <worker_name> --kind codex --pane <pane-id> -- --model gpt-6-luna -c model_reasoning_effort=high --yolo --no-alt-screen herdr --session <herdr_session> agent prompt <worker_name> "<task>" --wait --timeout 120000 herdr --session <herdr_session> agent get <worker_name> herdr --session <herdr_session> agent read <worker_name> --source recent-unwrapped --lines 120 # If still working, wait for the same turn; do not resend it. herdr --session <herdr_session> agent wait <worker_name> --timeout 120000 ``` Inspect the reply and task evidence: `idle`, `done`, or a timeout is not proof of success. ## Prompts and human handoff Write delegated task prompts directly in first person, preserving scope and limits. Keep routing metadata outside the worker prompt; omit “Ian wants” and delegation framing. When authentication may arise, include: > If you need me to enter a password, approve MFA, solve a CAPTCHA, or complete another human-only authentication step, stop and tell me what is needed. I can take over through Jump Desktop. Never request, read, type, copy, expose, or store my credentials. Stop at native app/site permission prompts too; never approve or bypass them. Report blockers to the origin. While blocked or an action is uncertain, retain the same runner thread, Codex process/conversation, and lock. Resume only after user steering and checking actual state; do not repeat completed actions. ## Complete and clean up When the task is verified complete and nothing is pending: 1. Reply to the origin with results, safe evidence, and any unverified claims; say cleanup is pending. Record completed actions and delivery in `owner.⁠md`. 2. Exit only this task's idle Codex, then verify the original shell is foreground, the task process is gone, and the dedicated agent list is empty: ```sh herdr --session <herdr_session> agent send-keys <worker_name> ctrl+d herdr --session <herdr_session> pane process-info --pane <pane-id> herdr --session <herdr_session> agent list ``` 3. Re-read `owner.⁠md`; require the token, request ID, and owner URL to match. Remove only that file, then `rmdir` the lock directory. Ownership drift or unexpected files stop cleanup. Preserve Herdr's session/workspace/pane and logs. 4. Report cleanup, then `update_thread(thread=<own URL>, archived=true)` only for this temporary runner thread, without worktree-removal flags. If reply delivery fails, leave the thread unarchived; do not replay the task. On failure or cancellation, establish the actual action outcome and confirm work has stopped before the same checked cleanup. If uncertain or cleanup fails, retain the lock and thread and report the blocker. A lock loser owns no resources and may archive its temporary thread after reporting the conflict. ```` config.json ```json { "runner_id": "ian-mac", "working_directory": "/Users/mac", "herdr_session": "amp-workers", "herdr_workspace_label": "amp-workers", "worker_name": "amp-gui", "owner_lock_directory": "/Users/mac/.local/state/amp-workers/owner.lock", "herdr_skill_path": "/Users/mac/.agents/skills/herdr/SKILL.⁠md", "codex_launch_args": ["--yolo", "--no-alt-screen"], "model": "gpt-6-luna", "reasoning_effort": "high" } ```

@theoSep 25

For those moving back to Claude Code after being on the ChatGPT/Codex app for awhile, how are you feeling about the move? What do you miss?

@rhyssullivanSep 25

Here's Rhys' practical guide to shipping an MCP your users want: - Your MCP should be able to do everything your dashboard can Yes including things like deleting resources. The way to make this practical inside of an agent while not having things be unsafe is to deep link them into your product from inside the MCP - Do not ship lazy loading / codemode inside of your MCP CodeMode is a harness detail. Most harnesses (Claude / Codex / OpenCode) support codemode now, the problem is if you have two codemode MCPs they don't compose nicely (Executor v2 will still support CodeMode, but, will start to default to a transparent tool proxy for better clients) - Ship a search docs / skills tool Agents love these and it helps give them more understanding of your product MCP is getting skills over mcp which will help here in the future - Have a deep link into product tool Let the agent generate a URL that brings them into your product, people still want to use your dashboard and data visualization! They just don't want to click through things - Customization I think this matters less over time, but, you can allow people to select 'toolsets' of things to connect with to clients. Some clients don't support customizing tool selections and then it's also nice to have tool permissions set to the auth token - Let people OAuth from whatever client they want Please stop putting restrictions on which clients can authenticate to MCP servers, your users hate it, it doesn't improve security, all it does is add more friction PostHog and Sentry are my two 'AI native' companies to look at today, they both have an excellent onboarding, MCP / CLI experience, if you need a place to learn from look at them To be honest though, while all the above is helpful just shipping your API spec + CIMD OAuth to it is enough for 99% of agents to have a good experience

@theoSep 24

It's insane how much better the $200 Claude Code plan is compared to Codex right now. Just a few weeks ago, it was the other way around. Excited to see how OpenAI fights back.

@aakashguptaSep 24

The product manager role is 3 jobs. Most people want 1. Here's how to hold them all: It's still the classic job that we saw when Martin Eriksson (@bfgmartin) drew it in 2011. You have to be at the intersection of what customers want, what will move the business forward, and what's just become recently possible. But with AI the PM has moved closer to everything. Customer What is worth solving? You want to get as close to the customer as possible. We used to be gated by UXR with access to our customers, with analysts with access to our data. All those barriers have dissolved, and with MCP or CLI connections, PMs are put right into the center of their most important customer data. And with AI automations, PMs can get more time to spend with actual customers. Tools: @hidovetail @pendoio @intercom @Zendesk @hotjar @Qualtrics Engineering What is possible? You still don't want to build THE thing (usually), but we're getting closer. We're building prototypes on the real code base and design system. We're skipping the old wireframe stage with Balsamiq and going straight into clickable, full-data prototypes to drive better discovery. Tools: @boltdotnew @lovable_dev @Replit @figma @github @vercel Business What is worth funding? You don't need to become the analyst, but you can't wait on one anymore. We used to file a ticket with the data team and wait a week for a dashboard. Now PMs query the warehouse directly, pull revenue by segment, and walk into planning with the business case already built. The number that decides the roadmap is one prompt away. Tools: @stripe @HubSpot @mixpanel @googleanalytics Looker @Snowflake The Product Manager Which one ships? You still make the call, but now everything is connected in one place. With agentic harnesses like Claude Code and Codex, you can set up an operating system at the center of everything. And with newer tools like Herdr and pi, you can really take the agentic part of that to a whole new level. You still make the call, but you have an AI employee helping you alongside everything. Tools: Claude Code (@claudeai), Codex (@OpenAI), @cursor_ai, @herdrdev, @pidotdev, Instinct (@noahrshinn) It's amazing how nothing changed. And everything changed.

@venturetwinsSep 24

Okay I think I've cracked the code on these. Seedance 2.5 is very good if you upload a reference video + images of new characters and ask to swap. In some cases the ref video will get rejected - so I just ask Codex to blur the faces and resubmit 😂 It's also best doing two characters at a time. For this one I did Dario and Jensen first and then re-ran it with Sam. Prompt below. Edit the entire source video @ Video1. Replace the viewer's FAR RIGHT performer with the man in @ Image1 in the same pink shirt and blue / gray hoodie. Keep the LEFT and MIDDLE performers the same. The photos provide identity and wardrobe only. The video provides motion, expressions, gestures, timing, interactions, camera cuts, framing, lighting and background. Preserve the original performance as faithfully as possible. Do not swap positions or invent new movement or scene elements.

@svpinoSep 24

Using managed agents is perfect for running coding agents anywhere in the world. The beauty of this: • You can resume a session from anywhere • You can fork a session into multiple agents • You can run parallel agents • You can reuse environments Here is how you can do this with Claude Code, Codex, OpenCode, or Hermes.

@mvanhornSep 24

Just successfully had Grok Bot ask Muse to make a phone call, all securely over Tailscale. Would anyone else want this tool I built for myself to let all my different agents talk to each other? Right now have it talking to Grok Bot, Instinct, Muse, Codex, Claude Code, Hermes and OpenClaw.

@rhyssullivanSep 23

there's a development flag in codex called: mcp_oauth_refresh_coordination you can enable it and i believe it'll fix the common MCP auth problem codex users face it's a development flag so yadadaya use at your own risk / ask your agent about it

@alliekmillerSep 23

For the last several years, I have said that OpenAI models are better at ideation, creation, and exploration. If I had to brainstorm something, I would have picked whatever OpenAI model was best in class. Yesterday, that changed. First, Opus 5.5 is the only model to hit (note: hit, not beat) my code-meets-poetry challenge. It is the only time I have shared the test with a lab. I was in shock. Second, last night when making slides, I prompted both Codex and Claude Code to give me design options for one slide I couldn’t crack. Codex was a mess. Claude Opus 5.5 made 6 versions as pngs using image 2.5 from OAI, labeled them perfectly, saved to computer, opened up finder, and messaged its preference (I agreed with one of its recommendations but not the other). When I said my two preferences, it immediately created them as editable slides in PPT (like text boxes, shapes, the whole thing). I didn’t ask for that. In my early testing of Opus 5.5, it had several tool call failures. And it seems they have fixed it. Opus 5.5 is a tool-wielding machine and feels far more creative than previous Claude models.

@thezviSep 23

okay, I'm not saying no one has to or that no one else will, but my Codex credits were just sitting there so I'm asking it to make the good version of the DataRepublican network graph, let's see what happens.

@mtsliveSep 23

.@eriktorenberg and @gaganbiyani explain why young people have more leverage than ever, when 1 person can now build what took 50 people a decade ago: Erik Torenberg: "When I was in college, we didn't have 18-year-old kids on X building these incredible things or even just contributing to the discourse. It's permissionless. You don't have to wait anymore to just build something incredible." "Because a lot of the technology is so new, it kinda resets the board. If you're not coming in with preconceived notions, in some ways you have an advantage." Gagan Biyani: "You can open up your laptop, use Replit, Claude, Codex, whatever, and build something that 50 people needed to build 10 years ago. But the world is also coming to you." "Most of the information you and I know is now literally in the public sphere. It's much easier for someone to come from the outside to enter the conversation and start contributing." @theacademysf

@mtsliveSep 22

SITUATION DETECTED: GPT-6 Sol and GPT-6 Luna are rolling out on ChatGPT, Codex, and API.

@danshipperSep 22

BREAKING: Anthropic just dropped Opus 5.5—and it’s pulling some of our recent Codex converts back to Claude. We’ve been testing it at @every across coding, design, writing, and knowledge work. It sometimes beats Fable 5.1 in our testing and is up to 40% cheaper than Opus 5. It's a strong contender for new daily driver model: It’s excellent at end-to-end builds that match your taste. I’ve started reaching for Opus over Fable on big, end-to-end coding projects. And @kieranklaassen is replacing Fable 5.1 with Opus for his day-to-day product work. He calls it his new favorite model. It produces legible prose, but still trails Astra on writing tasks. Opus 5.5 scores a 68.42 on reading ease—the highest on any model we've tested. But in my writing benchmark tests it consistently buries the main point in intro paragraphs, and revisions. You'll be able to understand what this model is saying (yay!) but for day-to-day writing it's still behind. The economics are striking. Anthropic says Opus 5.5 will cost $5 per million input tokens and $20 per million output tokens. That’s the same input price and 20% less for output than Opus 5’s $5/$25. Altogether it should save roughly 40% on costs than Opus 5. It still has a “do the most” problem. @hammermt tested it on our standard knowledge work benchmarks, and it's results were great when thye came back. On at least one test it ran past the 10 minute time limit before delivering. Net Result: If you build apps and interfaces, try it. It’s become my go-to for ambitious coding projects. I’m still roughly 80/20 Codex versus Claude in day-to-day use, and I still prefer Sol and Astra for editing. But I’m spending far more of my tokens with Claude than I was a week ago. State of Play: Codex is still the better harness for me, but Anthropic is steadily gaining ground. They have a history of making their smaller models perform better than their bigger ones (Sonnet 3.7 for example) and they seem to have done the same with Opus 5.5. Full vibe check will be on @every soon!

@aakashguptaSep 22

A single production agent still takes six vendors and six invoices. That's my read on why agent counts stay so low inside most companies. Building a useful agent got easy this year, and running one in production is where most teams still stall. The harness comes from one project, the sandbox from a compute vendor, tokens from an inference provider, storage from a cloud, a frontier lab for the hard steps, tools from a search API, and someone writes the glue holding it together. By month end the cost of one run is anyone's guess. DigitalOcean's answer is Managed Agents, out today in public preview, and the pitch is that the whole stack lives on one infrastructure and one bill. Claude Code, Codex, OpenCode, Hermes, LangGraph or your own container image run as they are, and the company says it handles the microVM, the inference, the storage and the tool access underneath. Two things caught my eye. Sessions are meant to outlive your laptop, so a multi-hour refactor keeps going after you close the lid and a teammate can pick it up elsewhere. And credentials get brokered at execution time, per the company, so the key stays away from the model and the sandbox. On the numbers, the claim is 16,000+ tools behind one gateway, pause and resume in 305 milliseconds, and CPU billed per second of actual use, so a paused agent sits at zero cost. Fourteen years ago a $5 Droplet made servers boring enough that people stopped counting them. Agents get boring the same way once running one is a single line on a bill, and that is when the count inside most companies climbs well past a handful.

@pmituSep 22

2026 is wild building an AI agent is the easy part now the real pain starts when you actually want it to do shit while you're away keeping sessions alive, managing sandboxes, connecting tools, paying 5 different providers... bro i just wanted my agent to finish the task 🤯 DigitalOcean just launched Managed Agents - bring your Claude Code, Codex or whatever agent you're using - they handle the runtime, tools and infrastructure You get to actually build stuff instead of babysitting your agents

@svpinoSep 22

Since I bought a separate computer to run my agents, my token usage and the things I’ve been able to automate have increased dramatically. Infrastructure is the number one bottleneck for agentic workflows. This here is even better: Run any agent with one command. Run one or multiple parallel instances. Preserve state between runs. And no complex setup necessary.