Loading ToolBench
Make several AI coding CLIs collaborate (and challenge each other) on ONE shared project. Exposes multiple AI agents (Claude, Codex, Antigravity, Grok) as MCP tools so they can delegate tasks to each other and run multi-agent relays.
Six tools with complete JSON schemas and descriptions, but significant quality gaps. Tool names lack clear action verbs (ask_claude, ask_codex, ask_agy, ask_grok are repetitive and vague about delegation semantics). Descriptions are adequate (100-150 chars) but generic, they repeat the same text across four nearly identical tools without explaining when to use each. Parameters are typed and described, but lack validation constraints (e.g., timeout ranges, enum values for 'auto' are present but not explained in descriptions). Output schemas are not documented, callers don't know what fields to expect from session IDs or relay transcripts. Error handling is absent: no guidance on retryability, failure modes, or recovery paths. The rainbow_relay tool combines multiple concerns (agent orchestration, mode selection, round management) without clear composition boundaries.
Delegate a task to Antigravity (a separate AI coding agent) and get its answer. It works in the shared directory, can read/edit files and run its own tools, and returns a session id you can pass back as `session` to continue the conversation. Use this to get a second model's implementation, review, or opinion.
Delegate a task to Claude (a separate AI coding agent) and get its answer. It works in the shared directory, can read/edit files and run its own tools, and returns a session id you can pass back as `session` to continue the conversation. Use this to get a second model's implementation, review, or opinion.
Delegate a task to Codex (a separate AI coding agent) and get its answer. It works in the shared directory, can read/edit files and run its own tools, and returns a session id you can pass back as `session` to continue the conversation. Use this to get a second model's implementation, review, or opinion.
Delegate a task to Grok (a separate AI coding agent) and get its answer. It works in the shared directory, can read/edit files and run its own tools, and returns a session id you can pass back as `session` to continue the conversation. Use this to get a second model's implementation, review, or opinion.
List which AI coding CLIs (claude, codex, agy, grok) are installed and available to delegate to.
Four nearly identical ask_* tools (ask_claude, ask_codex, ask_agy, ask_grok) with identical parameter schemas and boilerplate descriptions. LLMs cannot distinguish when to use each, they all claim to 'delegate a task' with identical signatures. This violates the single-responsibility principle and forces the LLM to guess which agent to invoke.
No output schemas documented. Callers don't know what fields session IDs contain, what structure the relay transcript has, or what fields list_ais returns. This forces LLMs to guess and makes chaining tools unreliable.
No error handling guidance. Tools accept timeouts and can fail (e.g., agent CLI not installed, network timeout, agent crash), but descriptions provide no recovery hints. LLMs don't know whether to retry, ask the user, or escalate.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
Run a full multi-agent collaboration: two or more AI coding agents take turns in one shared directory, keep their own memory, see each other's file edits, and challenge each other until the goal is met. Returns the full turn-by-turn transcript. Use for hard tasks worth throwing multiple models at.
Parameter descriptions lack validation constraints. 'auto' enum is defined (full|safe) but descriptions don't explain the difference or consequences. 'timeout' is an integer with no min/max bounds, LLMs could pass 0 or 999999. 'rounds' and 'agree' lack range guidance.
Tool descriptions are generic and repetitive. All four ask_* tools use identical boilerplate ('Delegate a task to [Agent]...returns a session id...'). Descriptions don't explain when to use Claude vs Codex vs Agy vs Grok, or what each agent specializes in. This forces LLMs to reason about agent capabilities without guidance.