Loading ToolBench
The open-source company brain that powers a closed loop — model-free, over MCP. Your AI tool (Claude, Cursor, Copilot) answers from your docs with citations or refuses; record how-work-gets-done as skills, govern real actions with human approval + an audit trail, and let the loop learn from outcomes.
Six tools with complete schemas and descriptions. Naming is clear and verb-driven (search_, ask_, ingest, list_, propose_, action_). Descriptions are substantive (100-250 chars) and explain WHEN to use each tool. However, parameter descriptions lack detail on constraints, formats, and error recovery. Output schemas are not documented. Error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint). The ingest and propose_action tools are write-heavy but lack confirmation/dry-run patterns. Overall solid foundation but missing production-grade polish.
Check the status of a proposed action (proposed | executed | rejected | failed) and its effect.
Ask the company brain a question and get a grounded, cited answer produced by the brain's own model. It cites the records it used and refuses if it has no grounding (it never invents).
Add knowledge to the company brain (text, CSV, or JSON). It becomes scoped, embedded, retrievable knowledge that every door can ground on, and triggers any configured fan-out agents.
List the provenance sources currently in the brain (within the allowed scopes), with how many records each holds.
Propose a real-world action (a notice, an update, a reply). The draft is GROUNDED in the brain (refused if no grounding) and enters the approval queue — a human (or rate-capped policy at L2) must approve before anything executes. Returns the action id to check status.
Search the company brain for grounded, access-scoped knowledge. Returns the most relevant records with their provenance (source) and similarity score, so YOUR agent can cite them. Use this to ground an answer before responding.
No output schemas documented. LLMs cannot predict response structure (fields, types, pagination). search_brain and ask_brain return text but no schema for structured fields like 'source', 'similarity_score', or 'citations'.
Missing tool annotations. ingest and propose_action are destructive writes but lack destructiveHint. search_brain and ask_brain are read-only but lack readOnlyHint. Agents cannot distinguish safe from risky tools.
Parameter descriptions lack format/constraint details. 'format' enum in ingest is present but 'scope' parameter has no validation guidance. 'scopes' in search_brain says 'comma-separated' but no example or validation rules. Agents cannot self-correct invalid input.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
No error recovery guidance. Tools return text responses but no structured error classification (retryable vs user-fixable vs fatal). If ask_brain has no grounding, the response is a plain string, agent cannot distinguish success from refusal.
propose_action is a write tool with no dry-run or confirmation step. Agents can propose actions that enter an approval queue, but no way to preview or cancel before submission. High risk of unintended side effects.