5dive Mcp

5dive MCP server

[npm] [Awesome MCP Servers]

Expose the 5dive agent-fleet CLI as Model Context Protocol tools. Point any MCP client (Claude Desktop, Cursor, Cline, or your own) at this stdio server to file tasks, inspect and message agents, and read the fleet digest — directly from inside a model context.

5dive is the CLI + control plane for running a fleet of autonomous coding agents as a self-governing company. This server is a thin, honest adapter: every tool shells out to the local 5dive binary's machine-readable --json surface and returns the result — so it inherits the CLI's auth, permissions, and audit log for free, and never handles secrets itself.

Tools

Requirements

• Node.js >= 18
• The 5dive CLI installed and on PATH (curl https://install.5dive.ai | sudo bash).

Install & run

npx @5dive/mcp        # run directly# ornpm i -g @5dive/mcp && 5dive-mcp

Client config (Claude Desktop / Cursor / Cline)

{  "mcpServers": {    "5dive": {      "command": "npx",      "args": ["-y", "@5dive/mcp"],      "env": { "FIVEDIVE_SUDO": "1" }    }  }}

Remote mode: 5dive-mcp serve (ChatGPT, OpenAI dots)

A client that runs in someone else's cloud can't spawn a stdio process on your box. serve puts the same six tools behind MCP's Streamable HTTP transport, on loopback, for your reverse proxy to publish over HTTPS.

Using managed 5dive? You don't need this: connect to https://api.5dive.com/mcp and sign in. See the guide.

# 1. Mint a token. It is printed once; only its hash is kept.5dive-mcp token create --name=chatgpt            # read-only5dive-mcp token create --name=chatgpt --write    # may also file tasks and message agents
# 2. Serve on loopback (default 127.0.0.1:8741, path /mcp).FIVEDIVE_SUDO=1 5dive-mcp serve
# 3. Publish it over HTTPS, e.g. with Caddy (automatic certificates):#    agents.example.com {#        reverse_proxy /mcp* 127.0.0.1:8741#    }

Then point the client at it:

• Clients that can send a header (Claude Code, the MCP inspector, most SDKs): https://agents.example.com/mcp with Authorization: Bearer <token>.
• ChatGPT developer mode offers only OAuth or "No authentication", so put the token in the URL and choose "No authentication": https://agents.example.com/mcp/<token>.

5dive-mcp token list shows the tokens; 5dive-mcp token revoke --name=chatgpt cuts one off, and a running server refuses it from the next request.

Security: what opening this endpoint means

• What a token can do. A read-only token can list agents and tasks, read a task, and read the digest. A --write token can also file tasks (task_create) and message agents (agent_send). No token can answer a gate, change settings, spend money, or run anything outside those six CLI calls, each built as a fixed argv with no shell. Whatever sender the remote client claims is dropped: its messages arrive labelled mcp-<token name>, and its tasks say in their body which connection filed them.
• Who it runs as. The tools run with the rights of the user running serve, plus root through sudo when FIVEDIVE_SUDO=1. The server only ever runs those six subcommands, but a compromised server process holds that user's rights, so run it as a dedicated user whose sudo grant covers only the 5dive binary.
• Rate limit. 60 calls a minute per token by default (--rate), counted per JSON-RPC message: batches are refused before anything runs. Bodies over 1 MiB are refused.
• Audit. Every tool call, allowed or refused, appends one line to audit.jsonl: time, token name, tool, outcome, client IP (from X-Forwarded-For when the proxy is on loopback), and the ids and names involved. Message text and task bodies are never written.
• Tokens at rest. tokens.json (mode 0600) holds a SHA-256 of each token, never the token. Both files live in $FIVEDIVE_MCP_HOME, default ~/.config/5dive-mcp.
• A token in the URL is a password in the URL. Your reverse proxy's access log records request paths. Turn path logging off for /mcp/*, or use header auth where the client supports it. Anyone holding the URL holds the token.
• Plain HTTP stays on loopback. serve refuses a public bind by default, because the token would cross the network unencrypted.

Prefer no inbound port at all? OpenAI's Secure MCP Tunnel runs a small client next to the server and connects out to OpenAI. Its tunnel-client can launch the stdio server directly (--mcp-command "npx -y @5dive/mcp"), but it needs an OpenAI Platform organization with tunnel permissions.

Configuration (env vars)

Safety

Arguments are passed to the CLI as an argv array with no shell, so tool input can never be interpreted as shell syntax. The server never sees secrets: the CLI reads its own credentials from the box.

Scope

This mirrors a curated slice of the CLI (tasks, agents, digest), not its full surface. It is a distribution and convenience layer, not a new API. For everything else, use the 5dive CLI directly (5dive --help). Full docs: https://5dive.ai/docs/5dive-cli

License

MIT