Problem
Captured context (screen stills, OCR markdown) is stored locally but inaccessible to AI assistants like Claude Desktop. There is no way for an AI to query "what was the user working on?".
Solution
Implement a stdio MCP server (src/mcp/) that exposes Familiar's SQLite queue and markdown files via the Model Context Protocol (JSON-RPC 2.0, protocol 2024-11-05).
Tools exposed
| Tool |
Description |
get_recent_context |
Last N minutes of captured markdown |
search_context |
Full-text keyword search over markdown |
get_session_context |
All captures from a specific session |
list_sessions |
Recent sessions with metadata |
Security
- Database opened read-only
markdown_path values boundary-checked to context folder before readFileSync
- No network exposure; stdio only
Acceptance criteria
Problem
Captured context (screen stills, OCR markdown) is stored locally but inaccessible to AI assistants like Claude Desktop. There is no way for an AI to query "what was the user working on?".
Solution
Implement a stdio MCP server (
src/mcp/) that exposes Familiar's SQLite queue and markdown files via the Model Context Protocol (JSON-RPC 2.0, protocol 2024-11-05).Tools exposed
get_recent_contextsearch_contextget_session_contextlist_sessionsSecurity
markdown_pathvalues boundary-checked to context folder beforereadFileSyncAcceptance criteria
src/mcp/server.jshandles MCP protocol lifecyclesrc/mcp/tools.jsimplements all 4 tools