Claude Code
Use the installer. It adds the server and the hooks, and the hooks are what make Claude Code show up reliably in Quaero and actually stop when stopped:
npx slaunt install --client claudeManual setup gets you the tools but no hooks, so the model has to call thread_begin/run_begin/run_end itself (install agent instructions so it does):
claude mcp add --transport http slaunt https://central-mcp-ccr-production.up.railway.app/mcpThen run /mcp, pick slaunt, and choose Authenticate.
What the hooks do
| Hook | Effect |
|---|---|
UserPromptSubmit | Binds the thread if needed, starts a run, captures the prompt |
PreToolUse | Calls tool_authorize; a deny blocks the tool call |
PostToolUse, SubagentStart, SubagentStop | Report activity, capture tool I/O |
Stop | Captures the final message and ends the run |
SessionStart (clear, compact) | Re-binds after /clear or compaction |
Claude Code fires the first SessionStart before MCP servers connect, which is why UserPromptSubmit handles initial binding.
Because the hook owns the run, a model-issued run_begin while it's RUNNING fails with -32602. Agent instructions written for hook-less clients can trigger this; it's harmless but noisy.
tool_authorize decisions
The PreToolUse guard is how stops take effect mid-turn. If a tool call is unexpectedly blocked, the reason tells you why:
| Reason | Decision | Meaning |
|---|---|---|
runtime_authorized | allow | Run is active and in scope |
runtime_stopped | deny | Stopped from Quaero or KillBar. The next prompt starts a fresh run. |
run_not_active | deny | No run is RUNNING, usually because UserPromptSubmit failed. Restart Claude Code. |
stale_turn | deny | Hook belongs to an earlier turn |
scope_mismatch | deny | Hook session doesn't match the authenticated connection |
invalid_metadata | deny | Hook payload is missing fields |