Loading ToolBench

Reading the current corpus…

mychatarchive MCP server · D · ToolBench

mychatarchive

Local-first AI memory archive. Import chat history, generate semantic embeddings, and search via MCP. Exposes tools for semantic search, recent messages, context retrieval, and thought capture across chat platforms.

<=2025-11-25STDIO5 tools

55/100

Definition Quality70

Protocol Readiness35

Supportability42

Spec posture

Inferred effective spec <=2025-11-25·current baseline 2026-07-28

Static source inference · medium confidence · detected: Logging

Deprecated protocol patterns detected

Summary

MyChatArchive provides 5 tools with good naming discipline (verb-first: get_, search_, capture_) and detailed parameter descriptions. However, output schemas are not documented in the provided source code, and error handling guidance is minimal. Tool descriptions are well-written (110-280 chars, within baseline p10=34, p90=392), and all parameters have type definitions and descriptions. The server uses FastMCP, a modern framework, but the implementation does not expose structured output schemas, making it harder for LLMs to plan downstream calls. Security is well-designed (sensitivity scoping, private content filtering), but schema documentation is the main gap preventing a higher score.

Tools (5)

capture_thoughtwritesource verified67/100

Capture a new thought, insight, or fact into the archive for future retrieval.

get_contextread onlysource verified72/100

Given a topic, return a comprehensive context bundle. Gathers related conversations, captured thoughts, and thread summaries to provide full context about a subject from your chat history.

get_current_datetimeread onlysource verified73/100

Return the current date and time (UTC). Use this when you need to know today's date, the current time, or temporal context for interpreting archived messages.

search_brainread onlysource verified72/100

Semantic search across all chat history by meaning. Finds messages most similar to the query using vector embeddings. Returns relevant messages with metadata (thread, timestamp, role).

search_recentread onlysource verified65/100

Retrieve recent conversations and captured thoughts by time range.

Issues (4)

HIGHresponse-shaper ↗

Output schemas not documented. Tool descriptions explain inputs but do not specify what fields the response contains. LLMs cannot reliably extract results or plan downstream calls without knowing the response structure (e.g., does search_brain return 'messages', 'results', 'chunks'? What fields are in each item?).

get_current_datetimesearch_brainsearch_recentget_contextcapture_thought

HIGHrecovery-guide ↗

Error handling lacks recovery guidance. Code returns JSON with 'error' key (e.g., invalid 'since' format) but does not suggest corrective actions. LLMs receive an error but have no guidance on what to try next or whether the error is retryable.

search_brainsearch_recentget_context

MEDIUMpaginated-result ↗

Result limiting not enforced or documented in descriptions. search_brain defaults limit=10, but descriptions do not warn that large limits could exhaust context. No explicit statement about pagination or result capping strategy.

search_brain

Recommendations

Document output schemas for all tools. Add a 'Returns' section to each tool docstring specifying JSON structure, field names, and types. Example for search_brain: 'Returns {"results": [{"message": str, "thread_id": str, "timestamp": str, "role": str, "relevance": float}, ...], "total": int}'.
Enhance error responses with recovery guidance. When 'since' format is invalid, return: '{"error": "Invalid since format. Use YYYY-MM-DD (e.g., 2024-01-15). Your input: \'2024/01/15\'"}'.
Add result limiting guidance to search_brain, search_recent, and get_context descriptions. Include: 'Results are limited to [limit] items by default. For large archives, use platform or group filters to narrow results.'
Document capture_thought idempotency. If using SQLite rowid or content hash for deduplication, state: 'Duplicate content (same tags and text) within 1 hour is treated as idempotent, retries return the same thought_id without creating duplicates.'
Consider adding pagination/cursor support to search tools. Add a 'next_cursor' field to responses and a 'cursor' parameter to allow LLMs to fetch large result sets incrementally without exploding context windows.
Add tool annotations (readOnlyHint/destructiveHint) to FastMCP registration. Mark search_* and get_* as readOnly=true, and capture_thought as destructive=false (it creates, not deletes). This signals to clients the safety profile.

Spec posture evidence

Inferred effective spec: <=2025-11-25.

Relies on Logging (deprecated) - log to stderr or use OpenTelemetry

Score history

Overall score trend

↑ 33 points across a rubric change (v1 → v2)

55/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-22	D	55	<=2025-11-25	v2
2026-03-09	F	22	-	v1