MCP Facts Server¶
sidekick mcp runs a local, read-only Model Context Protocol server over standard input/output. It lets the coding agent inspect the same provider-neutral quota, context, burn-rate, task, decision, note, and project-context facts shown by Sidekick.
Register the server¶
Install the CLI globally first (npm install -g sidekick-agent-hub), then register the sidekick binary from the project where you want to use it.
Use claude mcp list or codex mcp list to verify registration. Both commands launch the server in the current project directory; pass Sidekick's global --project <path> or --provider <id> options before mcp when an explicit target is needed.
Available tools¶
| Tool | Fact returned |
|---|---|
get_quota_status |
Current provider quota/rate-limit windows and reset times |
get_burn_rate |
Active session token burn-rate metrics |
get_context_pressure |
Context utilization, attribution, and compaction facts |
get_tasks |
Persisted tasks for the current project |
get_decisions |
Persisted architectural decisions |
get_notes |
Persisted knowledge notes |
get_project_context |
Composed tasks, decisions, notes, handoff, and project statistics |
Every tool is annotated as read-only, non-destructive, and idempotent. The server exposes no capture or store-mutation tools. get_quota_status calls the same shared resolver as sidekick quota: Codex always queries the live API for utilization and reset credits, falling back to the newer session or cached sample on failure (cache wins ties). Claude and z.ai may reuse a persisted sample under five minutes old. Results retain sample provenance, freshness, and any API failure. Dashboard refresh schedules are unchanged; the remaining facts come from local session and persistence data.
Example prompt after registration:
How much of my current quota window remains, and is my session close to context compaction?