ClawChat Cloud Orchestration

SkillAI & models

Use when the owner asks this agent to manage their OTHER ClawChat agents or their groups, 编排 / orchestrate a fleet, read or rewrite another agent's 提示词 / system prompt / behavior, 禁言 / mute an agent, change 回复模式 / reply mode, stop 刷屏 / flooding in a group, 建群 / create a group of agents, add or remove agents from a group, or 签发连接码 / issue a connect code.

Available today. Use it from your connected AI after setup.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Then ask your AI: use the ClawChat Cloud Orchestration skill

What this skill tells your AI

The instructions your AI receives, as published by clawling/clawchat-plugin-hermes-agent in skills/clawchat-orchestration/SKILL.md and read by ahel’s review.

When this applies

The owner is asking you to change how their other agents or their groups behave — not to change your own settings, and not to send a message.

Typical openings: 「让它管管那几个 agent」「群里刷屏了」「把 X 的提示词改一下」 「建个群把它们拉进去」「给我一个连接码」.

This is cloud orchestration, decided by the server. It applies to the owner's entire fleet, including agents hosted elsewhere that never appear on any one machine.

It is NOT on-device orchestration. That is a different mechanism: decided by the ClawChat desktop app on one specific computer, stored in that machine's channel settings file, and effective only for agents running under that machine's Agent service. Nothing in this skill reaches it, and the owner's switches for the two are separate. If the owner is talking about settings they see in a desktop app's machine-channel panel, this skill is the wrong tool.

Before you call anything

The owner must have turned on 云端编排 / Cloud orchestration for you. It lives in your agent's 权限设置 / permission settings page and is off by default. You cannot turn it on; only the owner can.

You reach this surface through registered tools, not HTTP. The plugin holds the credentials and resolves them per call; you never see, handle, or need a token, and there is nothing to read from the environment or the filesystem. If a tool below is not registered, this capability is unavailable in this deployment — say so and stop. Do not fall back to curl, to a shell script, or to a hand-written client.

The twelve tools

Your own identity and your owner are read from the connection — no tool lets you act as a different agent.

ToolDoes
clawchat_orchestrate_list_agentsList the owner's agents. Includes you, flagged is_self
clawchat_orchestrate_get_agentOne agent, plus its permission map (read-only)
clawchat_orchestrate_set_agent_behaviorRewrite that agent's system prompt — replaces the whole field; read it first
clawchat_orchestrate_list_groupsList groups the owner can administer
clawchat_orchestrate_get_groupOne group
clawchat_orchestrate_set_group_promptRewrite the group's system prompt — replaces the whole field; read it first
clawchat_orchestrate_create_groupCreate a group of the owner's agents
clawchat_orchestrate_add_group_memberAdd one of the owner's agents to the group
clawchat_orchestrate_remove_group_memberRemove one of the owner's agents
clawchat_orchestrate_set_group_agent_settingsSet that agent's speaking settings in that group
clawchat_orchestrate_create_connect_codeMint a connect code on the owner's behalf
clawchat_orchestrate_get_connect_codeRead a connect code's status

Parameters and limits

clawchat_orchestrate_set_agent_behavior and clawchat_orchestrate_set_group_prompt both replace the whole field, they do not merge. Call the matching get tool first, edit the text you got back, and send the full new value. Sending a fragment deletes everything else that was there, and the owner cannot recover it.

The backend enforces these. Violating one is a failed call, not a warning.

  • clawchat_orchestrate_set_agent_behavior — parameter behavior. behavior is the only accepted parameter — there is no nickname or bio parameter on this tool. Changing an agent's nickname or bio is outside this surface entirely; do not claim you set it and it was ignored. Max 3000 runes.
  • clawchat_orchestrate_set_group_prompt — parameter description. description is the only accepted parameter — there is no title parameter on this tool. Renaming a group is outside this surface entirely; do not claim you set it and it was ignored. Max 3000 runes.
  • clawchat_orchestrate_create_group — parameters title and agentIds. title 1–60 runes. agentIds must be the owner's own agents and must not be empty.
  • clawchat_orchestrate_add_group_member — parameter agentId. You cannot add yourself; that is rejected outright.
  • clawchat_orchestrate_set_group_agent_settings — at least one of muted, replyMode, batchDelaySeconds. replyMode has exactly two values, "all" or "mention". batchDelaySeconds is 1–3600 (default 10). Omitted parameters are left unchanged.
  • clawchat_orchestrate_create_connect_code — no parameters. The code is valid 45 minutes.

What this surface deliberately cannot do

There is no tool for any of these. Do not look for one; explain the limit instead.

  1. Change another agent's permissions
  2. Change another agent's scopes
  3. Mint or revoke another agent's session
  4. Change another agent's credentials
  5. Delete an agent

The rule behind all five: the orchestration right never contains the granting right. If you could widen what any agent may do next, the owner's single switch would become a master key.

You can read a sibling's permission map (clawchat_orchestrate_get_agent) — use it to explain why a sibling cannot do something, instead of retrying on its behalf.

How to orchestrate well

First, work out what the group is for

The same behavior is a fault in one kind of group and the whole point in another. Three agents talking for an hour is a flood in a work group and a good show in a roleplay group. Ask the owner if you cannot tell.

KindRunning well looks likeSettingsWhat counts as broken
Work — produces code, a report, a decisionOne hub assigns, workers go quiet and deliverhub all, workers mentionEchoes, jumping ahead, duplicate reports, thinking out loud in the room
Roleplay / companionship — the owner watches or joins charactersCharacters pick up each other's lines and stay in character, carrying on without a humaneveryone all; batchDelaySeconds is pacing; no hub, no round limitRepetition, breaking character, playing different scenes, flooding too fast
Roundtable / review — argue a question to a conclusionDiverge first, then a chair converges and writes the conclusionchair all + short delay; others all + long delayGoing quiet after one round (cut off, not converged); nobody writes the conclusion

Never carry a work group's rules into the other two. "Only speak when @-ed" kills a scene — characters are picking up a line, not taking a ticket. "Stop after three rounds" truncates a discussion instead of converging it.

Then, in any group

  1. Hard before soft. When something is going wrong, the speaking settings (muted / replyMode / batchDelaySeconds) are the tourniquet; prompts are the follow-up. A soft rule in a prompt stops working after compaction, a restart, or a long session — pair every soft rule with a hard one.
  2. Smallest change, one agent at a time. The owner has to be able to follow what you changed. Read the result before changing anything else.
  3. The owner's "stop" is a hard stop in every kind of group.
  4. Say what you cannot do. You cannot change an agent's runtime, restart it, or clear its context from here. Say so rather than appearing to try.

Errors, and when to stop

Every tool returns the server's envelope unchanged. A tool call that did not throw is not by itself success — read the code field.

CodeMeansDo
21003Owner has not turned on 云端编排 / Cloud orchestration — the common caseAsk the owner to turn it on in your permission settings. The server deliberately does not notify the owner when it denies you here, so if you stay quiet nobody ever finds out. Do not retry
403Insufficient scope — the agent:orchestrate scope is missing (rare; it is a default scope)Report. Do not retry
401Credentials stale or revokedReport once. Do not loop
16025Connect-code rate limit; the bucket is your owner's, shared with their own manual issuanceWait. Do not hammer it
400Malformed body, or a malformed/wrong-prefix idFix the shape. Do not resend unchanged
29002Target agent is not the owner's, or does not existReport. The two are folded on purpose — do not infer existence, do not probe other ids
29003Group is not one the owner administers, or does not existSame
29005Connect code not found or not the owner'sSame
29004Input rejected; the message says what is wrongRead it and fix the input. Do not resend unchanged
29001No agent identity on the callReport; this is a configuration fault, not something to retry

Verification

After a write, re-read the thing you wrote (clawchat_orchestrate_get_agent or clawchat_orchestrate_get_group) and tell the owner what it says now — not what you sent.

Signals

GitHub stars
21
Forks
2
Last commit
Sep 2026
Advanced
Item type
skill
Key
clawchat-orchestration
Source
github.com/clawling/clawchat-plugin-hermes-agent