MCP Server

The bridge between your checkpoint data and AI tools. When Claude Code asks "what was I working on?", the MCP server answers by reading from your local SQLite database.

What it is

The MCP server implements the Model Context Protocol, which lets AI coding tools discover and call structured tools at runtime. KeepGoing registers as an MCP server so tools like Claude Code can call get_reentry_briefing or save_checkpoint the same way they call any other tool.

The server runs as a subprocess, communicating over stdio. It starts on demand when your AI tool needs it - no persistent background process required.

Exposed tools

get_reentry_briefing

Returns a full re-entry briefing for the current project: last checkpoint, recent commits, session history. This is what populates Claude Code's project context at session start.

save_checkpoint

Saves a checkpoint with summary, next step, and optional blocker. Called automatically by the Claude Code hooks after each Stop event.

get_current_task

Returns the current active task and session info. Used by hooks and session-awareness features to track what's in progress.

get_decisions

Returns detected architectural decisions for the current project. Requires Decision Detection (Pro).

get_momentum

Returns momentum score and session data across projects. Used by the statusline and tray.

get_session_history

Returns recent session logs with timestamps and summaries.

get_whats_hot

Cross-project momentum summary sorted by recent activity.

get_context_snapshot

Returns a compact JSON snapshot of current project context for handoff to other AI tools.

CLI delegation

When the MCP server starts, it checks whether a keepgoing CLI binary is available on PATH. If it is, tool calls are delegated to the CLI instead of handled in-process. This means the CLI version drives behavior - update the CLI to get new features without updating the MCP server.

If no CLI is found, the MCP server falls back to its own in-process handlers. All functionality is available either way, but the CLI path is preferred for correctness and schema compatibility.

Supported clients