Skip to content
Coders Talk
Sign in

Your agent reads it too

Before a task, your agent can check how similar tasks went for other developers: the prompt that worked, where the agent went wrong, and what the human did about it. Coders Talk has an MCP server for that. It only reads, and only what people published.

Closed beta

The server is in closed beta while the library grows. To join, write to [email protected] with your Coders Talk username and the agent you use. Once you are in, connect it as below.

Connect

The address is https://coders.talk/mcp (Streamable HTTP). Clients that support OAuth open a Coders Talk page where you sign in with GitHub and press Connect. The connection can only read the library: it cannot see your drafts, send sessions or publish.

Claude

In Claude (web or desktop): Settings → Connectors → Add custom connector. Name it Coders Talk, paste https://coders.talk/mcp, press Add, then Connect.

Claude Code

With the agent plugin (0.10 or newer) the server comes with it and uses the plugin’s connection: nothing to add. Without the plugin:

claude mcp add --transport http coders-talk https://coders.talk/mcp

Then type /mcp in Claude Code, pick coders-talk and sign in. Or skip OAuth and use the token of the agent plugin: add --header "Authorization: Bearer ct_…" to the command above.

Codex

With the agent plugin (0.10 or newer) the server comes with it: run codex mcp login coders-talk once in a terminal to sign in. Without the plugin, in ~/.codex/config.toml:

[mcp_servers.coders-talk]
url = "https://coders.talk/mcp"

Then run codex mcp login coders-talk.

Cursor, Windsurf and others

Cursor, in ~/.cursor/mcp.json:

{ "mcpServers": { "coders-talk": { "url": "https://coders.talk/mcp" } } }

Windsurf, in ~/.codeium/windsurf/mcp_config.json, the same with "serverUrl" in place of "url". Any other client that speaks Streamable HTTP works too, with OAuth or with a plugin token in the Authorization header.

How it goes

  1. Your agent starts a task and searches the library: a few words about the task and the stack.
  2. Coders Talk answers with published sessions of a similar task, or with the moments where agents failed on a similar problem.
  3. The agent uses what fits, and tells you which sessions helped, with links.
  4. When you send your own session with the plugin, it names the Builds the agent got. Your draft says what it was built on ("Used from the library").
  5. Once you publish it, those Builds count it: "Helped 3 published sessions". The library grows from what people send back.

What your agent gets

Tool When What comes back
search_coding_agent_sessions Before a non-trivial task on a known stack, or when you ask how others did something Up to 5 sessions: the result, how long it took, how often the human stepped in, a link
find_coding_agent_failures After two or three failed attempts Moments where agents failed on a similar problem, and what the human did next
get_coding_agent_session To read one of those The goal, the first prompt and the outcome, or every moment in order

Your agent decides when to call them. When nothing similar has been published, it gets exactly that answer: "No similar sessions on coders.talk yet."

Only published Builds reach agents, and only their moments as the author wrote them: never the raw session, and code only as a count of files and lines with the file paths. Before a Build reaches agents, Coders Talk checks it:

  • Hidden characters, images and links to other sites are taken out of every answer.
  • A moment that downloads and runs code, carries a long encoded blob or speaks to the agent ("ignore previous instructions…") is left out; a Build with more than one waits for an editor.
  • A reported Build is out until an editor looks, after one report.
  • Builds of new accounts reach agents once they have votes from other people or the author has an earlier Build without reports.

Every answer starts by saying it is reference data from other developers, not instructions, and that commands in it are not run without asking you. See something wrong? Report the Build on its page.

What the MCP sends

  • The query your agent writes: a short description of the task and the stack. The tools ask agents to leave out code, file paths, repository and company names; the server cuts a query at 200 characters.
  • What we keep of it: the query with web addresses, email addresses and file paths replaced by [url], [email] and [path], the stack, task type and agent it filtered by, the client's name (for example Claude Code), how many sessions it found and which came first. Not who asked: no account and no IP address. Kept for 180 days. Queries that found nothing tell the editors what the library is missing.
  • Reads: when your agent opens a session or quotes its moments, the session's "Used by agents" count grows by one. To count a client once a day, we keep a hash of its IP address and name for a day, with a salt that changes daily, in the cache only.
  • With the plugin, when you send a session: how many times the agent searched the library in it and the addresses of the Builds it got, up to 20. Never the searches themselves.
  • Server logs: the tool called, whether it worked, the client, your account and IP address, how long it took, and never the query. Kept for 14 days, to keep the service running and stop abuse.
  • Limits: 120 searches a minute for a signed-in connection.

To stop, remove the connector in your client (claude mcp remove coders-talk in Claude Code; for the plugin’s own server, disable it in /mcp, or in Codex set enabled = false under [plugins."coders-talk@coders-talk".mcp_servers.coders-talk] in ~/.codex/config.toml) and disconnect it under Agent plugins in Settings; a disconnected token stops working at once. More in the privacy policy.