Connect your coding agent (MCP server)

Step 1 — add the server

Claude Code

claude mcp add --scope user --transport http coderbuds https://coderbuds.com/mcp/insights

Paste straight into a terminal. --scope user makes it available in every project — see managing the connection if you want it scoped differently.

OpenAI Codex

codex mcp add coderbuds --url https://coderbuds.com/mcp/insights

This adds Coderbuds to your Codex configuration without storing an API token. In step 2, run codex mcp login coderbuds to authorize it in your browser.

Cursor

Add this to ~/.cursor/mcp.json (global) or .cursor/mcp.json in the repo, then save:

{
  "mcpServers": {
    "coderbuds": {
      "type": "http",
      "url": "https://coderbuds.com/mcp/insights"
    }
  }
}

Everyone on the team at once

Commit that same block as .mcp.json in the repo root. Nothing secret is committed — each person authorizes once in their own browser the first time their client connects.

Step 2 — authorize in the browser

Your client opens a Coderbuds tab the first time it connects. Sign in if you are not already, click Authorize, and the tab hands control back to your editor. That grant is per person and per machine, so it happens once here and never again on this laptop.

If no tab opens, trigger it by hand: run codex mcp login coderbuds in Codex; run /mcp in a Claude Code session and choose Authenticate; or in Cursor open Settings → MCP and use the login control on that server. Leave your editor running while you authorize — it is listening on a local port for the redirect.

Step 3 — check it worked

Ask your agent a question only Coderbuds can answer:

A real answer means you are connected. If the tools are missing entirely, the client has not authorized yet — go back to step 2.

If you are on more than one team

Coderbuds never guesses which team a call is about. On one team, there is nothing to do. On several, either is fine:

• Per call (no reconnect): pass the repository you are working in, as owner/name. Coderbuds maps it to the owning team. Your agent can read that from git remote get-url origin.
• For the whole session: pin the team in the URL — https://coderbuds.com/mcp/insights?team=Your%20Team. Changing this counts as a different server, so your client will ask you to authorize once more.

Untracked repos — an internal tool, a scratch repo — will not resolve to a team on their own, so pin the team in the URL for those.

Managing the connection

Codex stores this server in ~/.codex/config.toml, shared by Codex CLI, the desktop app, and the IDE extension on this computer. Use codex mcp list to check it and codex mcp remove coderbuds to remove it.

Claude Code stores an MCP server in one of three scopes. Which one you used decides where the config lives, who else gets it, and how you remove it later. If you omit --scope you get local — not project, and not user.

Removing uses the same scope you added with:

claude mcp list                        # what is configured now
claude mcp remove coderbuds --scope user

After adding or removing, restart Claude Code — the server list is read once at session start, so claude mcp list inside a running session can report a server that is no longer configured on disk.

If something goes wrong

Every failure below names its own fix on screen too — this table is just the short version.

Fallback: headless environments (CI, servers)

Most people should not use this. Browser authorization above is the default and needs no token. Reach for a bearer token only where a browser cannot open — CI jobs, remote build boxes. Pass --header "Authorization: Bearer <your-token>" to the command in step 1, or an equivalent headers block in .mcp.json with the token supplied via an environment variable. A token is tied to you personally, so never commit one or share it with the team.