Coders Talk for Claude Code and Codex

Send the Claude Code or Codex session you are in to Coders Talk as a draft Build: the prompts, interventions and fails that mattered, with the raw log one layer down. The plugin never publishes. The draft is private: only you see it, or your team when the session ran in one of the team's repositories. You review every moment on the site and publish it there if you want to.

The other way round, the plugin gives your agent the Coders Talk library: before a non-trivial task, or after a few failed attempts, it can look up sessions where other developers did something similar, and what went wrong for them (see The library in your agent).

Install

From a terminal

curl -fsSL https://coders.talk/install.sh | sh        # macOS and Linux
irm https://coders.talk/install.ps1 | iex              # Windows, in PowerShell
coders-talk login

This installs coders-talk, one file with no Node.js needed, into ~/.coders-talk/bin and adds that folder to your PATH (the rc file of your shell; on Windows the user PATH). The installer takes the file for your platform from this repository's releases and checks it against the release's SHA256SUMS. No sudo or administrator rights. CODERS_TALK_VERSION=0.11.0 installs that version, CODERS_TALK_NO_MODIFY_PATH=1 leaves PATH alone.

coders-talk login waits in the terminal until you press Connect in the browser; over SSH it prints the link and the code to open elsewhere. coders-talk update replaces the file with the latest release after checking its SHA256. A command you type mentions a newer release in one line, at most once a day; it never waits for the network to do so, and hooks never check. CODERS_TALK_NO_UPDATE_CHECK=1 turns that off.

Then coders-talk enable connects Claude Code and Codex to it. It says what it found and what it will do, asks, and then installs the plugin through each agent's own plugin system: it lays the plugin out in ~/.coders-talk/plugin and adds that folder as the local marketplace coders-talk-local (claude plugin marketplace add, claude plugin install coders-talk@coders-talk-local; codex plugin marketplace add, codex plugin add). It is the same plugin as the one below, but its hooks and skills call coders-talk by its absolute path, so no Node.js is needed and desktop apps find it without your shell's PATH. On the way it:

• replaces the plugin from the GitHub marketplace (coders-talk@coders-talk), whose hooks would otherwise run twice;
• finds a Coders Talk MCP server you added by hand (claude mcp add, mcp_servers in Codex's config.toml) and removes it, or keeps it and installs the plugin without its own server for that agent;
• asks once about auto mode for every agent it found (off unless you say otherwise). Codex runs the hooks only after you trust them in /hooks; enable does not do that for you;
• in a git repository, asks whether to put the Coders Talk git hooks into it (no, unless they are there already; see Git hooks).

coders-talk enable --yes goes ahead without asking, with --agent=claude-code,codex, --auto=off|on|team|push, --mcp=remove|keep and --git-hooks, --no-git-hooks or --no-trailers for the answers. coders-talk status shows the file, the sign-in, each agent's Coders Talk plugins and auto mode, and the git hooks of the repository you are in. coders-talk disable uninstalls the plugin and its marketplace from both agents and takes the git hooks out of every repository enable put them in; the sign-in, your settings and the file stay. coders-talk update lays the plugin out again in the new version and updates it in the agents, so the plugin's version is always the file's.

To send a session without going back into it, run coders-talk sessions in the folder it ran in: it lists that folder's Claude Code and Codex sessions (also those run at the top of its repository), newest first, with how many prompts each has, the start of the first one and whether it was sent already. coders-talk build 2 then shows the same summary and privacy check as /coders-talk:build, asks [y/N] and sends; coders-talk build alone takes the newest session, and a session id works too. It takes the same options (--private, --team=<slug>, --keep=<numbers>, --continues=<link>). It runs only in a terminal: neither an agent nor a script can answer its question, and there is no --yes. From a script, coders-talk preview <id> and coders-talk send <id> (with --agent=codex for Codex) do the two steps.

A Coders Talk connector your organisation added in claude.ai is not visible on your computer, so enable cannot find it. If the agent then lists the library's tools twice, turn one of the two servers off in /mcp.

From the agent

1. In Claude Code:

   /plugin marketplace add coders-talk/coders-talk-plugin
   /plugin install coders-talk@coders-talk
   

2. Connect it to your account: /coders-talk:login. It opens coders.talk in the browser; check that the page shows the same code as Claude Code and press Connect. The plugin receives its token directly from the site and keeps it in ~/.coders-talk/credentials.json. There is no token to copy, and it never passes through the chat.

Needs Node.js 20 or newer, nothing else.

Codex

codex plugin marketplace add coders-talk/coders-talk-plugin
codex plugin add coders-talk@coders-talk

Start a new session, then $coders-talk:login and $coders-talk:build (the same commands as below, with $ instead of /). For the library, also run codex mcp login coders-talk once in a terminal. Codex asks to run the plugin's command outside its sandbox: it needs the network to reach coders.talk and your home folder for the sign-in. The session is found through CODEX_THREAD_ID in ~/.codex/sessions (or CODEX_HOME), and HEAD at its start comes from the rollout's session_meta, so Codex needs no hook for it. Codex runs the plugin's hooks only for auto mode (see Auto mode), and only once you trust them in /hooks.

One repository serves both agents: Claude Code reads .claude-plugin/ and skills/, Codex reads .codex-plugin/, .agents/plugins/marketplace.json, codex/skills/ and codex/hooks.json (its manifest points there, so Codex never picks up hooks/hooks.json; the same scripts run with --agent=codex). Codex does not expand ${CLAUDE_PLUGIN_ROOT} or ${CLAUDE_SESSION_ID} in skills, so its skills give the script path relative to the skill file and pass --agent=codex.

Use

Command	What it does
/coders-talk:build	Shows what would be sent (project folder, prompts, time span, size), where it goes and what the privacy check found, asks, sends, prints the draft link.
/coders-talk:build --private	The same, and the draft stays yours even in a team's repository.
/coders-talk:build --team <slug>	The same, and the draft goes to that team (the part after /t/ in the team's link).
/coders-talk:auto on	From now on every session of this agent on this computer is sent by itself while it runs and when it ends (see Auto mode).
/coders-talk:auto team	The same, only for sessions in repositories of teams that ask for it.
/coders-talk:auto off	Stops it. /coders-talk:auto alone says whether it is on and what it last sent.
/coders-talk:login	Connects this computer through the browser, or says which account it is connected to.
/coders-talk:logout	Forgets the token on this computer (revoke it on the site under Settings → Agent plugins).
/coders-talk:lookup <task>	Searches the Coders Talk library for sessions of a similar task and shows what came back. The agent also does this by itself (see below).

Sending the same session again updates its draft until you publish it.

Two rules:

• Commands that send your session (build and auto, and login and logout with them) run only when you type them. The agent never invokes them on its own.
• Searching the library is something the agent does by itself when a task calls for it (the lookup skill and the coders-talk MCP server). It sends a short description of the task and the stack, never the session. Switch it off in /mcp (Claude Code) or in ~/.codex/config.toml (Codex); see The library in your agent.

A forked session (/branch or --fork-session in Claude Code, a fork in Codex) is sent as its own draft, and the plugin tells the site which session it came from and where it left it. The fork's Build then says it is a fork and links to the Build of the original session, and the original's Build links to its forks. Whichever of the two is sent first, the link appears once both are on the site. Claude Code writes the lines the fork inherited with the original's session id and Codex names it in the rollout's session_meta, so no hook is needed for it.

If the privacy check found something the preview lists it by number; it goes as [REDACTED] unless you name its number, and the preview runs again with --keep=<numbers> (see What leaves your machine).

If coders.talk cannot be reached or answers with a server error, the send step opens coders.talk/new in the browser and the file manager with the prepared .jsonl.gz selected, and prints the link and the path: drop the file on the page. It is the same checked file the preview described. It is kept for 30 minutes, so sending again once the site is back works too. CODERS_TALK_NO_BROWSER=1 keeps the browser and the file manager closed; the link and the path are printed anyway. A rejected token or request never ends up there: sign in again instead.

Auto mode

Off until you turn it on, per computer, per site and per agent: /coders-talk:auto on in Claude Code never sends Codex sessions, and $coders-talk:auto on in Codex never sends Claude Code ones. With it on, each session is sent the way /coders-talk:build would send it, without asking: to your team's space when the repository is one of your team's, else to your private Builds. With /coders-talk:auto team, only sessions in repositories of teams that ask for it (a team setting) are sent, and nothing else leaves the machine. A team can ask; it can never switch this on for you.

Three hooks do the sending, each in a background process so the agent never waits for an upload:

• Stop, after an answer of the agent: when the session grew and the last send is ten minutes old, it is sent as still going. The draft stays up to date, and a crash loses at most those minutes. A session shorter than ten minutes is only sent at its end.
• SessionEnd: the session is sent once more, as ended. Codex also ends a session after 30 idle minutes, which is what ends one in its desktop app.
• SessionStart: sessions that never said they ended (a crash, a closed terminal) and grew since their last send are sent at the next start, up to three at a time. One quiet for half an hour goes as ended; a fresher one as still going. Only sessions auto mode saw while it was on are considered, never older history.

The site asks the model for moments once per session, when it is over: when the plugin says it ended, or after half an hour without anything new. Sending a session that is still going costs nothing and does not count against the daily limit.

Nothing is published by it. A session you already published is left alone. Every send gets a line in ~/.coders-talk/auto.log: sent or synced (with the draft link), skipped and why, or failed. ~/.coders-talk/auto-sessions.json remembers which sessions auto mode saw and how much of each went, never their content; /coders-talk:auto off forgets it. CODERS_TALK_AUTO=0 in the environment turns auto mode off for that shell.

In Codex the hooks are codex/hooks.json. Codex runs a plugin's hooks only once you trust them: type /hooks and trust the three Coders Talk hooks ($coders-talk:auto on reminds you). On Windows Codex runs them through PowerShell and puts the plugin folder into ${PLUGIN_ROOT} itself. When the Codex CLI exits it also ends what its hooks started, so the session-end hook waits up to three seconds for the upload; what does not make it goes at the next start, and the site asks for moments after half an hour without anything new anyway.

Private and team drafts

A draft is private: it is in your My builds, not in the feed, search engines or your profile. If you are in a team on Coders Talk and the session ran in a repository of one of the GitHub organisations the team named, the draft goes to the team's space instead: the team sees it, nobody else. The preview says where the draft goes before anything is sent; --private and --team <slug> override it. Publishing to the community is always a separate step on the site, and for a team's draft only if the team allows it.

The library in your agent

The plugin brings an MCP server, coders-talk (https://coders.talk/mcp), with three read-only tools: search_coding_agent_sessions finds published sessions of a similar task, get_coding_agent_session reads one of them, find_coding_agent_failures finds where agents failed on a similar problem and what the human did. The lookup skill tells the agent when to use them (before a non-trivial task on a known stack, after two or three failed attempts, when you ask how others did something), what may go into a query and what may not, and to treat the answers as other people's experience: it never runs a command from them without asking you. When a session helped, it says so with the link.

• Signing in. In Claude Code the server uses the sign-in of /coders-talk:login: .mcp.json asks scripts/coders-talk.mjs mcp-headers for the token at each connection (headersHelper), so the token is in no config file and never passes through the chat. Not signed in, the server asks you to sign in through the browser in /mcp. Codex runs a plugin's header helper from the session's folder with an empty environment, where the plugin cannot find its own script, so in Codex the server signs in on its own: run codex mcp login coders-talk once. Both need access to the library: during the closed beta, see coders.talk/for-agents.

• What a search sends. The query (a few words about the task, or the symptom of a failure), the stack, and the agent's name. The agent is told never to put code, paths, repository, company or client names, hostnames, URLs or secrets in it. The site keeps the query text, the filters and how many results it found for 180 days, not tied to you or your IP address. The session itself never goes with it.

• Switching it off. Claude Code: /mcp, choose coders-talk, disable. Codex: in ~/.codex/config.toml

  [plugins."coders-talk@coders-talk".mcp_servers.coders-talk]
  enabled = false
  

• Another site. The plugin's server always points at https://coders.talk/mcp: the Claude desktop app compares the address it shows with the session's and does not expand variables, so a ${…} address breaks signing in there. For another site add your own server and switch the plugin's off: in Claude Code claude mcp add --transport http coders-talk-dev http://…/mcp, in Codex [mcp_servers.coders-talk-dev] with url = "http://…/mcp" in config.toml. Sending sessions still follows CODERS_TALK_URL and the plugin option url.

• Which Builds your session used. When you send a session, the plugin counts the agent's calls to these tools in it and collects the Builds their answers linked to (/b/<slug>?ref=agent), up to 20; the preview shows them. The site links your draft to those Builds: "Used from the library" on yours, "Helped N published sessions" on theirs, once yours is published. Only the count and the slugs are sent, never the queries.

• A suggestion to share. Once per session, after an answer, when the agent got Builds from the library and the session changed code, the Stop hook shows you one line: "Your agent used 2 Builds from coders.talk in this session. Share yours: /coders-talk:build". It sends nothing, and it never shows with auto mode on. To read the session for it, the hook goes on from where it stopped the last time and keeps that place, the count and the slugs in ~/.coders-talk/nudges/. Turn it off with CODERS_TALK_NUDGE=0, or node <plugin folder>/scripts/coders-talk.mjs nudge off. In Codex the hook runs only once you trust the plugin's hooks in /hooks.

Git hooks

Only in a repository where you said yes to coders-talk enable (or --git-hooks); never globally. Two hooks, each a block between # >>> coders-talk and # <<< coders-talk right after the first line, so a hook already there keeps its own lines and disable leaves it as it was:

• prepare-commit-msg adds Agent-Session: <session id> to a commit that holds a session's work: a Claude Code session whose answers changed a staged file that has not been committed since (the git snapshots tell), or a Codex session that auto mode follows in this repository and that was written to since the last commit. Merges and squashes are left alone, and the same trailer is never added twice. --no-trailers leaves this hook out. The session id becomes part of the repository's history: in a public repository everyone sees it.
• pre-push finds the sessions behind the commits you push, by their trailers and by the Claude Code snapshots that saw those commits. With auto mode on, team or push, it sends them in the background, as ended. Without auto mode, it prints one line saying how many sessions are behind the push. The push never waits for the network and never fails because of Coders Talk, and the snapshot refs (refs/coders-talk/…) are never pushed.

Auto mode push sends a session only when its commits are pushed; sessions whose code you never push stay on your computer. The agents' own hooks send nothing in this mode.

A repository whose hooks live elsewhere (core.hooksPath: husky, lefthook, a shared folder) gets no hooks from enable: it prints the lines to add to those hooks yourself.

What leaves your machine

• By default, only the session you run /coders-talk:build in, and only after you confirm.
• With auto mode on, which only you can switch on, each session of that agent on this computer, without asking (with team, only sessions in repositories of teams that ask for it; with push, only sessions whose commits you push). Nothing is published either way.
• With the git hooks in a repository, the session id goes into your commits as an Agent-Session trailer, so it leaves with every push of those commits, to wherever you push them. The session itself does not.
• The session .jsonl without screenshots, thinking blocks and the agent's bookkeeping lines, with tool output cut to 60 lines, gzipped. The plugin's own run is cut off the end. Big files are read line by line: a 417 MB Codex rollout full of screenshots goes as about 200 KB.
• If the session ran in a git repository: the origin address when it is on GitHub, the branch, and the titles, hashes and size (files, +/−) of the commits the session made (the site links each commit on GitHub; the commits themselves are not sent). The address of a private repository is shown only to the draft's own audience (you, or your team) and is removed when the Build is published. In Claude Code a SessionStart hook remembers HEAD at the start of each session in ~/.coders-talk/sessions/ for this; it sends nothing and prints nothing.
• The files that changed, as diffs (up to 300 lines a file; env, key and credential files, lock files and builds by name only). In Claude Code scripts/snapshot.mjs runs on SessionStart, UserPromptSubmit and Stop and takes a git snapshot of the working tree: a temporary index in ~/.coders-talk/snapshots/, tracked changes and new files up to 1 MB, a tree chained under refs/coders-talk/<session>. Your branch, index and files are not touched, and a plain git push does not send that ref. The build step turns neighbouring snapshots into what the agent changed in each turn (Bash and subagents included), what you changed by hand between its answers, and the commits of each turn. A snapshot slower than 3 seconds turns them off for the session. Snapshots and their refs older than 14 days are removed at the next session start in the same repository. Without git, or in Codex, the diffs come from the agent’s own edits in the session.
• Nothing secret: keys, tokens, connection strings, passwords in URLs, email addresses, public IP addresses and internal hostnames are replaced with [REDACTED:TYPE] on this computer, before anything is sent, in the session, the branch name and the commit titles alike. A home folder in a path (C:\Users\you, /Users/you, /home/you) becomes ~. The preview lists each finding by number and kind, never its value; --keep=<numbers> sends one as it is, and only a SHA-256 of it goes to the site, so the site's own second check leaves it alone. Kept values are remembered as hashes in ~/.coders-talk/kept.json. Words to hide in every session, such as client names or internal services, go in ~/.coders-talk/privacy.json as {"redact": ["Globex", "billing-core"]}. The rules are the site's own (scripts/lib/privacy.mjs, generated from coders.talk), checked against the same cases in test/fixtures/privacy.
• The number of tokens the session spent per model (input, output, cache reads and writes), counted from the session file before it is slimmed. Only the counts; they show on the Build and in your team's numbers.
• For a forked session, the id of the session it was forked from and the time of the fork, so the site can link the two Builds.
• If the agent used the Coders Talk library in the session: how many times it called it, and the slugs of the Builds it got (up to 20).
• Apart from sessions: the library searches the agent makes by itself (a short task description and the stack; see The library in your agent).
• The token can start imports and read the library. Revoke it in Settings at any time.

Configuration

Setting	Where	Default
Site address	CODERS_TALK_URL, or in Claude Code the plugin option url (asked when the plugin is enabled). In Codex set the variable in ~/.codex/config.toml: [shell_environment_policy] set = { CODERS_TALK_URL = "https://…" }	https://coders.talk
Token	/coders-talk:login saves one per site; CODERS_TALK_TOKEN overrides it (for CI, or a token made in Settings by hand)	none
Data folder	CODERS_TALK_HOME	~/.coders-talk
Suggestion to share a session that used the library	CODERS_TALK_NUDGE=0 turns it off for a shell; coders-talk.mjs nudge off for this computer	on
Proxy	HTTPS_PROXY (HTTP_PROXY for an http:// site, ALL_PROXY for both), NO_PROXY for hosts that go direct. Lower-case names work too. An http:// or https:// proxy, with user:password@ if it asks; a socks:// one is ignored. In Codex, if a variable does not reach the plugin, add it to set as above	none: direct

Node's own fetch ignores the proxy variables, so the plugin opens the proxy's CONNECT tunnel itself. Some networks reset a direct connection to the site after the first 16 KB, which lets whoami through but not an upload; there the proxy is what gets a session out.

The script reads the url option from Claude Code's own settings.json (pluginConfigs). It does not rely on ${user_config.url} in the skills or on CLAUDE_PLUGIN_OPTION_* variables: in testing, the desktop app left the placeholder unexpanded and the variable did not reach commands the model runs.

Development