Loading ToolBench
HTTP + WebSocket daemon for Chrome extension. Acts as a bridge between AI agents and corporate systems via a Chrome extension that manages authentication and skill execution.
CorpGateway MCP exhibits solid naming conventions (verb-prefixed tools like cgw_list, cgw_invoke) and comprehensive parameter descriptions. Tool descriptions are well-written and contextual (avg ~150 chars). However, output schemas are entirely undocumented, no response field definitions, types, or structures are visible in the source. Parameter schemas are present and typed, but several tools lack critical constraint documentation (e.g., cgw_invoke's 'params' object accepts arbitrary string key-value pairs with no validation guidance). Error handling is minimal, no recovery guidance, categorization, or actionable error messages visible. Security concerns: confirmCode parameter in cgw_invoke suggests OTP handling but no validation or secret-injection pattern documented.
Permanently delete a knowledge note. Use sparingly — outdated facts are usually better updated than deleted, so a future cgw_recall does not surface stale info.
List available skill groups with descriptions. Call first to discover what corporate systems are connected.
Call a corporate skill. Only for skills where confirm=false in cgw_schema. If confirm=true, use cgw_invoke_confirmed instead.
Call a corporate skill that requires confirmation (confirm=true in cgw_schema). Use this instead of cgw_invoke when the skill has confirm=true.
List available skills. Returns compact text with skill signatures and descriptions. Optionally filter by group.
Search the knowledge base for notes (facts, skill annotations, recipes) relevant to a query. Use this when you need details that were not in the digest shown by cgw_list — for instance to look up org-specific naming, skill quirks, or a previously saved multi-step recipe. Returns top matches with full body.
No output schemas documented for any tool. LLMs cannot predict response structure, field types, or chaining IDs. cgw_invoke, cgw_list, cgw_recall responses are completely opaque.
cgw_invoke and cgw_invoke_confirmed accept 'params' as additionalProperties with no validation constraints, format guidance, or error recovery hints. LLMs cannot know what values are valid or what happens on invalid input.
No error handling guidance visible. Tools do not document retryable vs fatal errors, recovery steps, or actionable error messages. Agents cannot self-correct on failure.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
Save a new note to the knowledge base so future agent sessions can recall it. Use when you learn something durable about the org or about a skill — naming conventions, response field meanings, multi-step procedures. Pick the right "kind": fact for general org knowledge, annotation when the note is about a specific skill (pass skillName), recipe for a named playbook.
Get skill details. Returns parameters, confirm flag, and the invoke tool to use (cgw_invoke or cgw_invoke_confirmed). Always call this before invoking a skill for the first time.
Modify an existing knowledge note. Pass the note id and any subset of fields to change. Useful when refining a fact you previously recorded.
confirmCode parameter in cgw_invoke suggests OTP/secret handling but no documentation on format, validation, or secure transmission. Secrets in tool parameters risk logging and exposure.
cgw_recall and cgw_remember lack pagination support (no limit default, no next_cursor, no total count). Large knowledge bases could return unbounded results, exhausting context.