Use Third-Party Agents

Third-party agents are how provider-native agents join Monad Mesh. They run inside a Monad session as team members: Monad Mesh supervises the runtime, scopes it to one conversation, and presents its activity as a continuous timeline. The provider still owns model behavior, authentication, provider session identity, and provider-owned approvals. A mesh can be made entirely of third-party agents. Monad Agent Runtime is the first-party alternative for a member, not a prerequisite for one. For the HTTP, stream, and cursor contracts, see third-party agent observation. To add a provider, see Author a third-party agent adapter.

Supported providers

Monad ships adapters for eight providers:

Provider ID	Product
`codex`	Codex
`claude-code`	Claude Code
`antigravity`	Antigravity
`gemini`	Gemini CLI
`qwen`	Qwen Code
`openclaw`	OpenClaw
`hermes`	Hermes
`monad`	Monad (another Monad runtime driven as a mesh member)

Third-party atom packs may register additional provider IDs.

Check what is configured

The provider’s own CLI has to be installed on this machine. Start from the roster:

monad mesh agents          # configured third-party agents, with their provider and state
monad mesh auth codex      # the provider's sign-in state

Sign-in is not brokered. A third-party agent is a separate product with its own credentials, so Monad never proxies its login: sign in with that provider’s CLI (codex login, claude, …) and Monad picks the session up. monad mesh auth only reports the state, and a provider that exposes no probe reports unknown. In the browser, Studio → Mesh agents shows the same roster and is the only surface that can run an interactive provider login inside an embedded terminal.

Run one in a session

A third-party agent always runs against a Monad session — the transcript it reports into. A project-bound conversation still has its own ses_… id; a project id is not accepted.

ID=$(monad session new "codex on this repo" --json | jq -r .id)

monad mesh start codex --session "$ID" --cwd .   # spawn the runtime
monad mesh list                                  # live runtimes, daemon-wide
monad mesh show mesh_123456789012                # lifecycle, activity, connection

mesh start refuses to spawn a runtime that would only sit at a login prompt, and rejects a --cwd outside the session’s configured project root.

Watch and drive it

monad mesh watch  mesh_123456789012              # follow the neutral projection
monad mesh watch  mesh_123456789012 --raw        # verbatim provider frames (diagnostic)
monad mesh input  mesh_123456789012 "add tests for the parser"
monad mesh steer  mesh_123456789012 "stop, do the migration first"
monad mesh interrupt mesh_123456789012
monad mesh stop   mesh_123456789012
monad mesh usage  [mesh_123456789012]            # provider overview or one runtime's tokens

watch prints one line per observation event and resumes from the last cursor after a dropped connection; --json emits NDJSON. --raw is a privileged diagnostic surface: it can contain prompts, tool arguments, file content, environment details, or credentials. Approvals raised by the provider are answered the same way:

monad mesh approve mesh_123456789012 <requestId> [--reason "…"]
monad mesh deny    mesh_123456789012 <requestId> [--reason "…"]

Whether a runtime accepts steering, interrupts, or approval resolution depends on its adapter — read the session’s effective capabilities rather than assuming, and see native CLI approvals for the autopilot switch. The web UI renders the same runtime as a live card in the transcript with the same controls; the CLI is the scriptable path and the only one usable headlessly.

Use reported capabilities

Agent and preset responses may include this capability object:

type MeshAgentCapabilities = {
  auth: 'pty' | 'status-probe' | 'none';
  events: 'paged' | 'provider-owned' | 'none';
  resume: 'pty' | 'structured' | 'none';
  approval: 'provider-owned';
  autopilot?: boolean;
  fastMode?: boolean;
  settingsImport?: boolean;
  approvalProxy?: boolean;
};

Use these fields to decide which setup and history controls to show. Session controls come from the effective runtime capabilities returned with each MeshSession. Do not infer capabilities from the provider name or expose provider process topology.

Build against the API

Every verb above is a thin wrapper over the daemon’s Mesh API — POST /v1/mesh/sessions to start a runtime, /stream/convenience and /stream/raw for the two observation planes, /events/convenience for history. A client written against it must resume with Last-Event-ID, treat the ready frame’s eventsBefore cursor as the history anchor, and replace local state when the server sends a replacement snapshot instead of a patch. The request and response shapes, cursor rules, and epoch semantics are specified in third-party agent observation. Observation never writes chat: only Message Ingress creates durable messages.

Common failures

Symptom	Meaning	Next step
The agent is missing from `monad mesh agents`	It is not configured	Configure it in Studio → Mesh agents, or import its settings with `monad import settings`
The preset reports `installed: false`	The provider executable was not found	Install the provider CLI, then re-check
`monad mesh auth <agent>` reports `unauthenticated`	Provider credentials are missing or expired	Sign in with that provider’s own CLI
`monad mesh start` rejects `--cwd`	The path is outside the session’s configured project root	Choose a path inside that root
A control returns unsupported capability	The active provider runtime cannot perform it	Use the session’s effective runtime capabilities
Convenience history is empty	The provider has no readable event source or projection	Check raw observation and adapter support
Stream replays from the epoch start	The resume cursor is malformed, foreign, or stale	Replace local live state with the new `ready` anchor
Raw page reports `settled` coverage	Provider history omits transient transport deltas	Treat it as settled history, not byte-complete live capture

Get Started

Monad Mesh: The Agent Team Runtime

Monad Agent Runtime: The First-Party Agent

Operate the Daemon

Trust and Policy

Reference

Use Third-Party Agents

Supported providers

Check what is configured

Run one in a session

Watch and drive it

Use reported capabilities

Build against the API

Common failures

​Supported providers

​Check what is configured

​Run one in a session

​Watch and drive it

​Use reported capabilities

​Build against the API

​Common failures

Supported providers

Check what is configured

Run one in a session

Watch and drive it

Use reported capabilities

Build against the API

Common failures