Exposes brain.db as structured MCP tools. Works with: Claude Code, Cursor, Windsurf, Codex, OpenCode, Gemini, etc.
ShipFast MCP server has 14 tools with schemas and descriptions present, but quality is inconsistent. Descriptions are present but often generic or vague (e.g., brain_linked: 'Show linked repos...' is 60 chars, below the 194-char baseline). No output schemas documented. Error handling is minimal, most tools return raw query results or empty arrays on failure, providing no recovery guidance. Tools like brain_decisions, brain_learnings, brain_context, brain_tasks, and brain_findings support write operations but lack confirmation/dry-run patterns. Parameter validation is weak (e.g., brain_trace accepts max_depth up to 10 with no bounds checking visible). No tool annotations (readOnlyHint/destructiveHint) despite clear read/write distinction. Composition is reasonable, tools are single-purpose, but chaining is not optimized (e.g., brain_search returns node IDs but downstream tools like brain_trace require specific 'fn:file:function' format, forcing manual ID construction).
Get or set context variables. Context stores project state, requirements, and configuration.
List or add decisions. Decisions are architectural choices, trade-offs, and reasoning stored in the brain.
List indexed files, optionally filtered by name or extension.
List or manage findings. Findings are research topics, summaries, and citations with hash-verified freshness.
Show most changed files in the codebase. Useful for identifying hotspots and areas of active development.
Walk the code graph to find impact of changes. Trace consumers (upstream) or dependencies (downstream) of a node.
List or add learnings. Learnings are patterns, problems, and solutions discovered during development.
Write operations (brain_decisions, brain_learnings, brain_context, brain_tasks, brain_findings) lack confirmation/dry-run patterns. Agents can destructively modify state without preview or rollback capability.
Most parameters lack descriptions. E.g., brain_decisions 'phase' param has no description; brain_trace 'kinds' param description is terse.
No output schemas documented. LLMs cannot predict response structure. E.g., brain_search returns 'matching nodes' but exact fields are not specified.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | 2026-07-28+ | v2 |
Show linked repos and their brain.db status. Use shipfast link to connect repos for cross-repo search.
Search files/functions by name or pattern. Returns matching nodes with file paths, signatures, and line numbers.
List skill execution sessions. Sessions track command runs, outcomes, and branch context.
Get brain.db statistics: node count, edge count, decisions, learnings, hot files, tasks.
Full status summary of the brain: node/edge counts, recent decisions, learnings, hot files, active tasks, and available MCP tools.
List or manage tasks. Tasks represent work items tracked in the brain.
Find paths between two nodes in the code graph. BFS through edges, returns every path up to max_depth. Answers questions like "how does X reach Y?" or "what call chain connects the button click to the db write?". Uses the same edges table as brain_impact.
Error handling is minimal. Tools return empty arrays or generic errors on failure (e.g., query() returns [] on exception). No recovery guidance for LLMs.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Despite clear read/write distinction in Risk field, MCP protocol annotations are absent.