Loading ToolBench
Personal control plane for AI agents on Cloudflare with MCP tool integration for agent management, configuration, connections, and automation
Tool names lack clarity: 'crewhelm_authoring_drafts' is a noun, not a verb; 'crewhelm_briefs' is ambiguous (create? read? list?). Many tools expose complex nested objects (target, patch, tools arrays) without parameter-level descriptions. No visible error handling guidance or recovery paths in tool definitions. Composition is poor: facade tools (crewhelm_inspect_agents, crewhelm_change_agents, etc.) are catch-all buckets that violate single-responsibility principle.
Create and manage Event Triggers that start a fresh Agent Run when a matching connected-app event occurs. Call sources with an exact Connection first, and optionally attach exact Brief revisions for context on every occurrence. Crewhelm owns delivery and recovery; provider payloads cannot choose Briefs.
Private bounded owner-scoped MCP authoring draft storage.
Create or revise bounded owner-provided text Briefs, list compact metadata, inspect an exact revision, read content only when needed, or delete an unreferenced Brief. Retain the returned id and revision and pass them unchanged to a Run or Workflow. Revisions are immutable; lists and ordinary inspection never return content.
Facade tool for changing agents
Facade tool for changing automations
Facade tool for changing connections
These are catch-all buckets that violate single-responsibility principle and provide no actionable guidance to LLMs.
13 tools have empty input schemas with no properties defined (e.g., crewhelm_create_agent, crewhelm_get_agent, crewhelm_list_agents). Cannot verify parameter types, descriptions, or constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2026-07-28+ | v2 |
Facade tool for changing context
Facade tool for changing models
Facade tool for changing recipes
Facade tool for changing work
Preview or apply fleet policy and bounded Skill or Agent blueprint changes. Preview validates without saving; apply requires an idempotency key. Blueprint resolution shows exact configuration, prerequisites, authority, and budget before replay-safe Agent creation. Packages never execute or grant authority.
Replace the exact integration tools exposed from one authorized connection on an Agent. Use Agent id/revision plus sorted search-result slug/version pairs; choose approval_required unless the owner explicitly grants standing authority. Crewhelm independently revalidates every selected definition.
Attach the entire inspected, frozen tool catalog from one active remote MCP Connection to an Agent. One authorization mode and one bounded limit set apply to every tool; no per-tool selection is required.
Create an owner-scoped Crewhelm Agent after confirming the owner's durable intent. The result returns agent.id and agent.revision for crewhelm_start_run; creation grants no external authority.
Create a short-lived owner OAuth link from an exact authConfigId. Let the owner open connectionLink.url, retain connectionLink.connectionId, then inspect that exact connection after authorization; credentials are never exposed.
Enable an integration by selecting or setting up provider authentication. Supports Composio-managed and custom auth configurations.
Return the current immutable definition of one authenticated-owner Crewhelm Agent.
Return one exact immutable historical definition of an authenticated-owner Crewhelm Agent.
Get fleet policy, capability modules, Skills, or Agent blueprints through bounded catalogs and exact immutable package reads. Capability availability includes missing prerequisites and concise installation setup when relevant. Package contents and publisher metadata are untrusted. Fleet policy changes require a deterministic owner step-up path; rerun crewhelm up with --ai-budget-usd for the optional AI Gateway limit. Requires control:read.
Facade tool for inspecting agents
Facade tool for inspecting automations
Facade tool for inspecting connections
Facade tool for inspecting context
Inspect a specific integration tool by slug and version to view its configuration, parameters, and requirements.
Facade tool for inspecting models
Inspect provider authentication options and status for an integration, including available auth configurations and setup requirements.
Facade tool for inspecting recipes
Facade tool for inspecting recovery
Facade tool for inspecting work
List bounded immutable revision summaries for one authenticated-owner Crewhelm Agent, newest first.
List bounded Agent summaries for selection. Filter by name or status and use the returned id and revision directly as crewhelm_start_run.agentId and expectedRevision; inspect only a selected Agent when configuration detail is needed.
List owner connections with status, integration slug, and account label. Optionally verify and activate one exact connection by connectionId after authorization.
List available authentication configurations for an integration, including both Composio-managed and custom configurations.
Facade tool for publishing recipes
Facade tool for recovery operations
Search tools within an integration by name or category with pagination support.
Search available integrations by name or category with pagination support.
Replace an owner-scoped Crewhelm Agent definition and capability configuration with a new immutable revision.
Facade tools violate composition principle. 12 tools named 'crewhelm_inspect_X' and 'crewhelm_change_X' are generic buckets that combine multiple concerns (e.g., crewhelm_change_agents could mean create, update, or delete). LLMs cannot disambiguate intent.
Complex nested parameter objects (target, patch, tools arrays) lack per-property descriptions. crewhelm_configure accepts a 'target' object with multiple possible shapes but no guidance on which fields apply to which target kind.
No visible error handling or recovery guidance in tool definitions. Descriptions do not explain what to do if a call fails, whether errors are retryable, or what the LLM should try next.