Loading ToolBench
Cross-tool memory for AI agents. Enables seamless context handoff between Cursor, Claude Code, Codex, Windsurf, Gemini, Copilot, Continue.dev, Zed, and Aider. Preserves task context, decisions, snapshots, and progress across agent sessions.
Stitch MCP server has well-structured tool definitions with clear naming (verb_noun pattern) and comprehensive descriptions. All 10 tools have input schemas with type definitions and required field declarations. However, output schemas are not documented, parameter descriptions lack depth (missing format/constraint details), and error handling guidance is absent. Tool descriptions are strong (100-200 chars, LLM-optimized), but parameter-level documentation is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Composition is sound, tools are single-responsibility and chainable via task_id references.
Record a technical decision. WHEN: Any architectural, design, or implementation choice where alternatives existed. WHAT: The problem, what you chose, what you rejected, and why. Future agents use this to avoid repeating failed experiments. Duplicate decisions (same problem) are auto-rejected.
Idempotent project bootstrap. Call this FIRST at the start of every session. It silently initializes .ahcp/ if needed, creates the active_task file, and ensures all directories exist. Safe to call multiple times.
Create a new task with a title and objective. This also sets it as the active task.
Read the full context.md file for a task — the human-readable living document with all context.
Generate a compact handoff bundle for a task. Contains objective, state, decisions, and recent activity — everything needed to resume work.
Get full details of a task including objective, state, decisions, and recent snapshots. Use 'active' as task_id to get the currently active task.
Output schemas not documented. Tools return data but LLMs cannot infer response structure (fields, types, pagination). Forces agents to guess what data is available.
Parameter descriptions lack format/constraint details. E.g., 'message' in stitch_snapshot says 'min 10 chars' in tool description but not in parameter schema. 'status' enum in stitch_update_task is not documented in parameter description.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Server declares risk classifications (READ_ONLY, WRITE) but does not expose them via MCP tool annotations, limiting client-side safety guardrails.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
List all tasks in the current project or globally. Use this to discover what tasks exist and find the active task.
Search for tasks by keyword across titles, objectives, tags, and decisions.
Save a snapshot of progress. WHEN: (1) After completing any sub-task or step, (2) After a failed experiment (prefix with 'FAILED:'), (3) Every 3-5 meaningful actions as a progress checkpoint, (4) Before session end. WHAT: Be specific — 'Implemented JWT auth with RS256, tokens expire in 1h' not 'worked on auth'. Duplicates are auto-rejected.
Update the current state, next steps, blockers, or status of a task.
Error handling guidance absent. No recovery hints in tool descriptions (e.g., 'If task not found, call stitch_list_tasks to discover available tasks'). Agents cannot self-correct on failures.
stitch_auto_setup has empty inputSchema properties object. While the tool is idempotent and safe, the schema should explicitly document that no parameters are required/accepted.