Loading ToolBench
Multi-agent reasoning MCP server for Claude Code. Spawns parallel research agents to find knowledge LLMs don't have, cross-validates against live docs and real code, and exports corrections as training data.
Symposium has well-structured tool definitions with clear descriptions and proper schemas. All 4 tools are explicitly registered with names, descriptions, and input schemas. Descriptions are detailed and include usage guidance (10 - 200 chars range). However, output schemas are not documented, LLMs cannot predict what fields to expect from tool responses. Parameter descriptions are present but some lack format/constraint details. Error handling guidance is minimal. The server demonstrates good intent but lacks production-grade polish in output documentation and error recovery patterns.
Multi-agent research engine that makes you smarter. Spawns parallel Nia-powered agents to research a coding task from multiple angles, then synthesizes findings into a verified plan. USE THIS TOOL WHEN: - You need to use an API or library you're not 100% sure about (it may have changed since your training data) - The user asks you to implement something involving a specific library version (e.g., 'add BetterAuth v3 session management') - You're about to write code that calls an external API and want to verify the current correct usage - You hit an error that suggests you used a deprecated or changed API - The user asks you to research how to implement something before coding it DO NOT USE when: the task is pure logic/algorithms with no external dependencies, or for simple file operations. The tool learns from every run. If it has researched a similar problem before, it returns instantly from stored knowledge. FEEDBACK LOOP: If the response includes `recalled_entry_ids`, call `symposium_feedback` AFTER the user confirms the implementation works (or after tests pass) to report whether the recalled knowledge was correct. This makes future recalls more accurate.
Export the Symposium knowledge base as structured training data. Returns all stored knowledge entries in a format suitable for fine-tuning: each entry has a 'mistake' (what models get wrong) and 'correct' (the verified right answer), with evidence sources. USE THIS TOOL WHEN: - The user asks to export or download the knowledge base - The user wants to see all discoveries in a machine-readable format - For generating fine-tuning data from Symposium's discoveries
Provide feedback on recalled knowledge to improve future accuracy. WHEN TO CALL: After the user confirms the implementation works, or after tests pass. Do NOT call immediately after generating code. Wait for verification. USE THIS TOOL WHEN: - You implemented code using Symposium's recalled knowledge and the user confirmed it works (outcome: correct) - The implementation failed or needed changes (outcome: incorrect) - It partially worked but needed adjustments (outcome: partial) - The recalled_entry_ids from a previous symposium call are available
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). symposium returns recalled_entry_ids but no schema for the full response object.
symposium_knowledge 'limit' parameter lacks min/max bounds. Unbounded numeric parameters invite LLMs to pass absurd values (e.g. limit=999999).
symposium 'depth' parameter has enum constraint but no guidance on when to use each level. LLMs cannot reason about cost/latency tradeoffs without hints.
Error handling is implicit. No recovery guidance in tool descriptions. If NIA_API_KEY is missing, the server returns a generic error, LLMs cannot self-correct without explicit guidance.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
List knowledge that Symposium has learned from previous research runs. Each entry represents a verified discovery: something that models typically get wrong, and the correct usage found from live documentation or real GitHub repos. USE THIS TOOL WHEN: - The user asks 'what has Symposium learned?' or 'what do you know about X library?' - You want to check if Symposium already has knowledge about a library before running full research - After a symposium run, to show the user what was stored
symposium_feedback 'detail' parameter is optional but no guidance on when it's required. Unclear if outcome='partial' demands detail or if it's always optional.