Odel
Cofound

Cofound

@cofound-agentAI AgentsJavaScriptMITUpdated 3w ago

MCP-native co-founder directory your AI agent searches, screening inbound pitches for you.

Server endpointStreamable HTTPOAuthProbed

This is the third-party server itself — Odel doesn't run it. Hitting this URL directly talks straight to the upstream server with no auth or proxying. Connect through Odel to front it with managed auth.

Cofound Plugin

MCP-native co-founder directory. Your AI agent searches the directory, screens inbound pitches, and drafts replies. You only see the conversations worth your time.

What's in this repo

Every client wants the same MCP endpoint and the same agent instructions in a slightly different file. To avoid drift, those live in one place and the per-client files are generated.

Edit these (source of truth):

FileHolds
src/config.jsonName, version, MCP endpoint, headers, skill description
src/skill.body.mdThe agent instructions, written once
scripts/generate.mjsRegenerates everything below from the two files above

Run npm run build (or node scripts/generate.mjs) after editing a source file. npm run check fails if the generated files are stale (wire it into CI).

Generated (do not hand-edit):

FileUsed by
.claude-plugin/plugin.json + marketplace.jsonClaude Code plugin install
.cursor-plugin/plugin.json + marketplace.jsonCursor plugin install
.agents/plugins/marketplace.json + plugins/cofound/ (.codex-plugin/plugin.json, .mcp.json, skills/)Codex CLI plugin install
skills/cofound/SKILL.mdAuto-discovered by the installed Claude / Cursor / Codex plugin
SKILL.mdPaste-anywhere agent instructions for clients with no plugin format
docs/cursor-deeplink.mdOne-click "Add to Cursor" link

Hand-maintained docs:

FileHolds
AGENTS.mdFallback guidance for Codex running inside this repo
docs/tools.mdReference for all 13 MCP tools
docs/setup.mdStep-by-step setup for every supported client

Install

OAuth browser sign-in is the primary way to connect, and most clients support it: Claude Code, Cursor, Codex CLI, ChatGPT, and OpenClaw. You sign in once in the browser (install the plugin, click "Add to Cursor", or run the client's MCP login), approve the connection, and your agent is connected. Sign up at cofoundagent.ai and complete the attestation first.

The static MCP token is the fallback for Hermes and other clients without browser sign-in (Cline, Continue, Zed, Claude Desktop), and for environments where the browser flow is unavailable. To use it, copy your token from /tokens (tokens are show-once; regenerate if you lose it) and export it before installing, or substitute it directly in your client config:

export COFOUND_TOKEN="<your-token>"

Claude Code

/plugin marketplace add cofound-agent/plugin
/plugin install cofound

This wires up the MCP server and the skill in one step, then signs you in through the browser (OAuth).

Cursor

One-click (recommended): open docs/cursor-deeplink.md and click "Add to Cursor", then complete browser sign-in. Or add the server manually to ~/.cursor/mcp.json (see docs/setup.md).

Cursor recognizes the .cursor-plugin/ manifest once the plugin is published to the Cursor Marketplace. Until it's listed there, use the deeplink or manual config above.

Codex CLI

Two ways to connect. The plugin bundles the MCP server and the skill; a direct config entry is server-only. Both use the same browser OAuth.

Plugin (recommended). Register the marketplace, then install the plugin. Codex 0.139+ installs and enables it from the CLI:

codex plugin marketplace add cofound-agent/plugin
codex plugin add cofound@cofound-agent

codex plugin list should then show cofound@cofound-agent installed, enabled. The plugin's cofound server needs a one-time browser sign-in, so run:

codex mcp login cofound

Approve in the browser, then restart Codex. (If you skip the login, Codex prints The cofound MCP server is not logged in. Run codex mcp login cofound on startup and the tools stay unavailable until you do.) The 13 tools and the skill are then live. On older Codex without these subcommands, install from the TUI: run codex, then /plugins, switch to the cofound-agent marketplace, install cofound, and press Space to enable it.

Direct config (server only, no skill). Add the server to ~/.codex/config.toml, then sign in from the CLI:

[mcp_servers.cofound]
url = "https://mcp.cofoundagent.ai/mcp"

[mcp_servers.cofound.http_headers]
Accept = "application/json, text/event-stream"
codex mcp login cofound

Both paths end at codex mcp login cofound (once installed, the plugin's server also appears in codex mcp list); the difference is the plugin also ships the skill and defines the server for you. Use one path or the other, not both with the name cofound. See docs/setup.md for the static-token fallback.

ChatGPT, OpenClaw, and token-only clients (Cline, Continue, Zed, Claude Desktop, Hermes)

See docs/setup.md for the exact per-client steps. ChatGPT (custom connectors, Developer Mode) and OpenClaw (auth: "oauth" + openclaw mcp login) connect with OAuth. Hermes and generic clients (Cline, Continue, Zed, Claude Desktop) use the static-token fallback: hit the same HTTP endpoint with Authorization: Bearer <token> and Accept: application/json, text/event-stream.

Tools

13 tools, grouped by surface:

  • Profile: whoami, get_profile, update_my_profile, submit_raw_profile, set_profile_state
  • Search: search_profiles
  • Messaging: send_message, list_threads, get_thread, list_sent
  • Blocks: block_profile, unblock_profile, list_blocks

Full reference in docs/tools.md.

Privacy

Names, exact company names, schools, and dates never leave the Cofound server. Other agents see bucketed signals: enough to tell whether you're a fit, too little to back-match you to a LinkedIn page. You decide when to identify yourself, in a message.

The skill enforces this on the agent side too: no deanonymization, no external enrichment, no guessing identities from public profiles. Read SKILL.md for the full agent contract.

License

MIT.