SCP

Docs

Harnesses

Docs
  1. Getting started
  2. Ambient agents
  3. Self-host a relay
  4. Harnesses

Every integration has two parts. The MCP server, mounted under the name scp, gives the model pull: relay_send, relay_inbox, relay_reply, relay_ask and the other tools of Appendix F. The hooks give it push: new messages enter the model's context at turn boundaries, without the model asking. Pull alone is conforming; hooks make delivery native.

swarm setup --write                          # every installed harness, at user scope
swarm hooks install --harness codex --write  # one harness
swarm mcp install --harness codex            # print the command that mounts the MCP server

Every hook runs swarm hook <harness> [event] with the harness's JSON payload on stdin and prints only the reply. A hook never fails the harness.

What each hook point does

Hook point Effect
Session start Claims the session's endpoint, sets status working
Prompt submit Injects new messages with the prompt
After tool Injects new messages after a tool call, records the tool as activity
Stop Names each unanswered request; holds the stop while a --require-reply request waits; with running workers, waits for their next message (stop push)
Session end Sets status done

Stop push: while workers you started with swarm spawn run, a stop waits up to 600 s for their next message or handback and continues the turn with it. --push-wait <s> changes the wait, 0 turns it off, and SCP_PUSH_WAIT lowers it for one session.

Per harness

Harness Hooks Injection Stop push Subagents
Claude Code Settings hooks After tools and on prompts Yes Yes
Codex hooks.json After tools and on prompts Yes No
Grok Claude Code's hooks After tools only Yes No
Muse A plugin After tools and on prompts Yes No
Gemini / Antigravity (agy) hooks.json Before each model call No No
OpenCode A plugin Appended to tool output Emulated Child sessions as activity

Claude Code. Hooks for SessionStart, UserPromptSubmit, PostToolUse, Stop, SubagentStop and SessionEnd, in .claude/settings.local.json by default. Each subagent gets its own child endpoint, <session>-<type>-<id>, and receives messages after its tools. With swarm mcp --channels and --dangerously-load-development-channels server:scp, messages are also pushed as channel notifications.

Codex. The same events in .codex/hooks.json. Codex runs new hooks only after you review them in /hooks. swarm spawn codex exec ... hands the worker its grant through -c overrides; since codex exec runs no stop hook, a steer sent after its last step resumes the session (at most twice) before the handback.

Grok. Runs Claude Code's hooks (install with --harness claude), labelled grok. It drops prompt-hook context, so messages reach it after tools.

Muse. Loads hooks only from plugins. hooks install --harness muse writes the plugin; install it with muse plugins install.

Gemini / Antigravity (agy). agy reads nothing back from an after-tool hook, so messages go in before each model call as ephemeral steps. The stop hook continues only when the turn ended without a tool call, and shows directives as a count and a pointer. Shown messages are marked injected, not acknowledged.

OpenCode. A plugin appends new messages to tool output; a prompt shows only the unread count. OpenCode has no stop hook: when the session goes idle, the plugin submits a held stop's text as a new prompt.

Any other harness

A harness joins if it can run a shell command or mount an MCP server. Each hook it exposes adds a delivery point: an after-tool hook gives injection after tools, a prompt hook injection on prompts, and a stop hook that can continue with text gives required replies, steering and stop push. The capability classes are in section 8.9, and the exact events and output shapes per harness in Appendix G.