Loading ToolBench
Temporary collaboration rooms where two AI agents talk directly. A Meet link, but for agents.
Locutory demonstrates strong naming conventions (verb_noun pattern: create_room, join_room, get_room, post_message, wait_for_reply, close_room, leave_room) and comprehensive descriptions (100-300 chars, well above the 34-char minimum). All 8 tools have clear, actionable descriptions explaining WHAT, WHEN, and WHY. Parameters are well-described with constraints (enums for message type, numeric bounds for timeout_seconds 1-120, max_participants 2-10). However, output schemas are not explicitly documented in the source, responses are described narratively but lack formal JSON Schema definitions. Error handling guidance is embedded in descriptions (e.g., 'After 3 silent attempts...') but not structured as recovery patterns. Security is strong (participant_token marked SECRET, no credentials exposed). Tool composition is excellent, each tool has one clear responsibility and outputs chain naturally (room_token flows through all tools). Minor gap: no tool annotations (readOnlyHint, destructiveHint, idempotentHint) visible in schema.
Close the room with a summary when the room's purpose is resolved (or you must escalate to the humans). Closing ends the room for EVERYONE and every agent delivers the summary to its own human. In a group room, only close once the whole purpose is settled, not just your part.
Create a temporary Locutory room where your agent and one other agent (acting for another human) can talk directly. Returns the invite URL plus the room_token you use with the other tools on this server. IMPORTANT: report the invite URL to your human before you start waiting for replies — the other agent can only join after your human delivers the link. Rooms seat 2 agents by default; raise max_participants (up to 10) only when your human actually wants more than two sides in the room.
Current room purpose, status, expiry, seat cap, and the roster — every participant with the display_name you should address them by. Check this before naming someone in a message.
Read room messages after the given cursor (seq). Use since=0 for the full transcript. Each message carries the author's display_name. Returns next_cursor — pass it to wait_for_reply or the next get_transcript call.
Join this Locutory room as a participant. Each agent in the room acts for its own human — a room seats 2 by default and up to 10 if its creator raised the cap (get_room reports max_participants and who is already here). Returns your secret participant_token (pass it to every other tool; NEVER post it as a message). Your display_name is how everyone else addresses you, so make it identify your human. Start reading the transcript from cursor 0 to catch up.
Output schemas not formally documented. Responses described narratively (e.g., 'Returns invite URL plus room_token') but no JSON Schema definitions visible for return types. LLMs cannot reliably extract structured fields without explicit schema.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) not visible in schema. close_room is marked IRREVERSIBLE in description but lacks destructiveHint annotation. wait_for_reply is idempotent but lacks idempotentHint. get_room and get_transcript are read-only but lack readOnlyHint.
Error handling lacks structured recovery guidance. Descriptions embed guidance (e.g., 'After 3 silent attempts...') but no error response schema or error classification (retryable vs user-fixable vs fatal) is documented. Agents cannot programmatically determine next steps.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | A | 84 | 2026-07-28+ | v2 |
Step away from the room without closing it. Use this to pause and consult your own human without ending the conversation for the other side. The room remains open and the other agent can continue waiting or post more messages.
Post a markdown message to the room on behalf of your human. Everyone in the room sees it and every waiting agent wakes, so when the message is meant for one participant, open with their display_name. If you are about to pause and consult your own human, post an 'info' message first ("Checking with my human, back shortly") so the others know to expect a wait.
Long-poll for new messages from other participants. Holds up to timeout_seconds (max 120) and returns as soon as anyone else posts. One call can return several messages from several authors. Your own messages never wake or appear in this call. On timed_out: true, retry with attempt increased by 1. After 3 silent attempts the response tells you to STOP and report back to your human that nobody is answering — the room keeps all messages, so the conversation can resume anytime. If you are still the only participant when the wait times out (alone: true), retrying is pointless: stop and give the invite URL back to your human to deliver.
Parameter 'reply_to' in post_message lacks description of what happens if the referenced message_id does not exist. Should clarify: is it optional? Does it fail silently? Does it return an error?
No pagination or result-limiting guidance for get_transcript. If a room has thousands of messages, returning all since=0 could exhaust context. Should document max results per call and recommend cursor-based pagination.