Claude Code Integration

Fastest path — once per machine:

nerveplane setup                                          # global hooks + login service + register this repo
claude mcp add --scope user nerveplane -- nerveplane mcp  # MCP server for all projects
# restart Claude Code

This installs the hooks at user scope (~/.claude), so there's no per-repo setup — every agent you launch auto-registers, and warnings/DMs are injected before edits.

If you'd rather wire a single repo (project scope):

claude mcp add nerveplane -- nerveplane mcp   # register the MCP server
nerveplane install claude-code                # project-scoped hooks + instructions
# restart Claude Code

1. The MCP server (claude mcp add)

claude mcp add nerveplane -- nerveplane mcp registers Nerveplane's seven MCP tools the idiomatic way. The spawned stdio server is a thin bridge that proxies to the daemon's REST API, keeping the daemon the single writer.

• Scope: add --scope user (all your projects) or --scope project (commit a shared .mcp.json); default is local to you.
• HTTP transport (no stdio bridge): claude mcp add --transport http nerveplane http://127.0.0.1:7734/mcp — the same seven tools are served over both transports from one definition.

2. The hooks (nerveplane install claude-code)

This is the piece claude mcp add can't do. Three hooks:

• SessionStart → runs nerveplane session-start, which auto-registers the agent for the current worktree the moment a session begins — so registration never depends on the model remembering to call register. It also seeds the session with a short coordination summary (active peers + "call sync").
• PreToolUse → routing puts a warning in an agent's inbox, but routed ≠ read — MCP has no reliable server→agent push. Before the agent runs Edit/Write/MultiEdit, this hook resolves the agent for the current worktree and injects any unread high-severity warnings and direct messages straight into its context. It always exits 0 and never blocks the edit.
• Stop → runs nerveplane stop-check when the agent finishes a turn. A Claude agent is turn-based with no background loop, so a teammate's DM would otherwise sit unread until the agent next acts. If there are unread direct messages, this hook returns a block decision ({"decision":"block","reason":…}) so the agent handles them before going idle — autonomous reply, no human needed. It's loop-safe (honors stop_hook_active) and only blocks on direct messages (not info events). This makes concurrently-active agents ping-pong on their own; waking a fully-parked agent needs the always-on worker model (see the roadmap).

nerveplane install claude-code writes (under .claude/ for the project, or ~/.claude/ with --global):

• settings.json — the SessionStart + PreToolUse + Stop hooks.
• nerveplane-instructions.md — the agent protocol (below).
• appends the instructions @import to CLAUDE.md (idempotent) so they load automatically — no copy-paste.

Flags: --global (user scope — install once, applies to all repos), --with-mcp (also write a project .mcp.json — use this if you don't have the claude CLI), --print (dry-run).

3. The agent protocol (auto-imported into CLAUDE.md)

1. You're auto-registered at session start; call register to add your capabilities/task and read the join packet.
2. Periodically and before finalizing, call sync.
3. Before changing API contracts, DB schemas, or shared types, call publish.
4. Record durable decisions with decision.
5. To talk to a specific agent, use chat (and chat action='wait' to block for a reply). When you finish a turn with unread teammate messages, the Stop hook asks you to handle them before going idle.

Other clients

The Streamable HTTP endpoint (http://127.0.0.1:7734/mcp) works with any HTTP-MCP client (Cursor, Codex, custom agents) — register it the same way with claude mcp add --transport http … or your client's MCP config.