cancel_run
ChatGPTCancel runId through POST /api/runs/:id/cancel, with Web ownership and cancellation side effects. Returns runId, status and message.
get_chat_activity_summary
ChatGPTRequest the existing Web public activity summary for threadId and runId.
get_chat_indicators
ChatGPTShow active (pending/running) and unread (completed/failed/cancelled) indicators for visible Agents and threads. This limited, non-atomic list does not prove Run completion. Use get_chat_messages with the known threadId and check terminal events for the matching runId to confirm an outcome. Does not mark threads read.
get_chat_input_status
ChatGPTRead-only, body-free observation of one original accepted input from retained Chat events. Does not look up execution records, download archives or paginate history. Successful JSON is at most 2 KiB. Use the MCP send eventId or native clientEventId. Native callers omitting clientEventId must obtain the original input ID once from raw history; createdAt alone is not unique. A retained replacement can establish the input's disposition even when its original row is unavailable. queued, rejected and revoked describe the input's retained disposition. consumed confirms consumption only; it does not confirm that execution is still active. completed, failed and cancelled require the same consuming runId's run.completed, run.failed or run.cancelled event in this thread. Multiple inputs may share that runId, which remains a correlation reference after execution-record deletion. terminal contains Chat event coordinates, not execution creation/start/completion times; Chat failures do not expose a separate timeout status. unavailable means retained metadata cannot establish the input; it never proves queued or completed. For archived evidence use the separate snapshot/raw-history operations.
get_chat_messages
ChatGPTGet one conversation's R2 snapshot pointer (null before compaction) and one bounded raw event-row page with canonical eventId, threadId and nullable runId references. Event identities, replacement references, paired cursors and payloads keep their native meaning. Use the known threadId to check recent sends without relying on search. Omit the cursor initially; rows start after the snapshot boundary, or at zero when snapshot is null. To inspect archived events, download snapshot.url (gzip NDJSON); downloaded rows retain native Web format, including id and chatThreadId. While hasMore is true, continue with cursor.lastSeqId as sinceSeqId and cursor.lastEventId as sinceEventId. On 410, restart from a fresh snapshot. Does not download, decompress or reconstruct history.
get_chat_thread_list
ChatGPTGet the current user's chat thread list in the authorized organization as an R2 snapshot pointer and one bounded page of lifecycle events, with canonical eventId and threadId references. Download the snapshot and apply lifecycle events to reconstruct the list; this tool does not return one individual thread or download or reconstruct the snapshot. Omit sinceSeqId initially; when hasMore is true, continue with the last event's seqId until all pages are read. Downloaded snapshots retain their native Web format.
list_agents
ChatGPTList agents using GET /api/agents. The JSON text is the unmodified Web response array.
list_models
ChatGPTList available models using GET /api/run-models.
rename_chat_thread
ChatGPTRename threadId with the existing Web rename body. Returns no content on Web 204.
revoke_queued_message
ChatGPTRecall the queued input identified by eventId through the ordinary Web recall event. Maps eventId to native revokesEventId; no original-input lookup or before/after state checks.
search_chat_messages
ChatGPTSearch using keyword, agentId and optional since. since is an inclusive ISO 8601 datetime lower bound (matchedMessage.createdAt >= since), with seconds, at most three fractional digits, and a required Z or ±HH:MM timezone, for example 2026-10-12T08:02:50Z. Numeric timestamps are not accepted. Selects up to 500 scoped keyword candidates without ordering, then sorts only those candidates by time and returns at most 100 messages. It does not guarantee the newest matches across all history. Pagination is not supported, and no continuation cursor or completeness signal is returned. Fewer than 100 results do not prove that all matches have been found. Hits and matchedMessage expose threadId; matchedMessage.runId is nullable. Hits expose stable agentId and nullable canonical displayName matching list_agents, for currently visible Agents. The index updates asynchronously: recent messages may be missing even after a Run completes, and an empty result does not prove absence. No index freshness signal or lag bound is returned. To confirm a recent send, use get_chat_messages with its known threadId and inspect the snapshot and event pages; do not resend solely because search is empty.
send_chat_message
ChatGPTSend plain text through POST /api/chat/events using Web validation and enqueue semantics. Generates and returns the accepted input's eventId with threadId and createdAt. Use eventId with revoke_queued_message to recall it while queued. Use get_chat_input_status with threadId and eventId for a bounded read of its retained Chat disposition and matching terminal event. Consumption alone does not prove completion. Does not wait for execution or provide safe retries after a lost acknowledgement.
update_chat_thread_model
ChatGPTUpdate threadId with the existing Web model-selection body, including its effort/tier/event semantics. Returns no content on Web 204.