---
title: Agents in Venom Chat
description: Add the Venom agent, or a service agent from Venom Tasks, to a channel and mention it; its answer streams into the thread. A channel owner picks its tools, including read access to the other Venom products, which run as you; a write waits for your Confirm.
section: cloud
order: 172
audience: [user, operator]
sources: [services/nodejs-cloud/lib/chat/work-model.ts, services/nodejs-cloud/lib/computers/sink-chat.ts, services/nodejs-cloud/lib/chat/agent.ts, services/nodejs-cloud/lib/chat/agent-model.ts, services/nodejs-cloud/components/chat/AgentSection.tsx, services/nodejs-cloud/app/api/cron/chat-agents/route.ts, services/nodejs-cloud/lib/chat/agent-products.ts, services/nodejs-cloud/lib/chat/agent-venom.ts, services/nodejs-cloud/migrations/0201_chat_agents.sql, services/nodejs-cloud/migrations/0205_chat_agent_tools.sql, services/nodejs-cloud/migrations/0270_chat_agent_members.sql, docs/adr/0059-venom-chat.md]
features: [cloud.chat]
---

The **Venom agent** is a member you add to a [Venom Chat](/docs/cloud/chat.md) channel. Mention it, and it reads the conversation,
works, and answers in the thread under your message.

## Add it to a channel

Open a channel's details and find **Venom agent**. A channel owner turns it on and ticks the tools it may use there:

| Tool | What it does | Runs as |
|---|---|---|
| Search this channel's messages | Finds earlier messages in this channel only. | the channel's agent member |
| Search the organisation's notes | Searches the notes shared with the whole organisation, never anyone's private notes. | the organisation |
| Search Venom Tasks issues, Read a Venom Tasks issue | Finds and reads issues. | the person who mentioned it |
| List Venom Incidents errors, List open Venom Incidents pages | Lists error issues (open by default), and who is on call with the open pages. | the person who mentioned it |
| List a Venom product's read tools, Read from a Venom product | Reads from People, Funnel, Support, Calendar, Flags, Incidents, Experiments, Computers, Notes, Tasks, Forms, Status and Org (see below). Never changes anything. | the person who mentioned it |
| Create a Venom Tasks issue, Comment on a Venom Tasks issue, Acknowledge a Venom Incidents page, Resolve a Venom Incidents error (**Writes**) | Proposes a change. | the person who mentioned it, after they confirm |
| Start work on a Venom Tasks issue (**Write**, service agents) | Proposes to take the request on: an issue made from your message, or the one it links, is handed to the agent, and the thread follows its progress and pull requests. See [From a channel message to a merged pull request](/docs/cloud/chat-work.md). | the person who mentioned it, after they confirm |

Above the tools there is a **Level**: **Can propose changes** (the default) or **Read-only**. A read-only agent is offered
no write tool at all, whatever is ticked, and a proposal made before the switch is refused when someone presses Confirm.
Changing only the tools keeps the level.

With no tool ticked it answers from the conversation alone. Anyone in the channel sees whether it is on. A direct
message cannot have the agent.

## Ask it

Write `@Venom` in a message (pick it from the mention list). A **Working…** reply appears in the thread at once; when the
job finishes, the reply becomes the answer and lists the tool steps it took ("Searched this channel: "deploy" (3 found)").
A mention in a thread answers with that thread as context; a mention in the channel uses the last 15 messages.

The answer **streams in**: the reply shows the text as the model writes it, then settles on the final answer and the tool
steps. The reply is updated about once a second, not for every word, so a long answer is a handful of edits. Stop
cancels the stream: the reply keeps "Stopped by …".

## Service agents from Venom Tasks

A **service agent** is a Venom Tasks agent an admin bound to a team key. Under **Service agents** in the channel details a
channel owner can add one to the channel, with its own tools and level. Mention it by name (`@Triage Bot`) and it
answers in the thread **as itself**: the reply is authored by it, it reads only this channel, and it uses only the tools
ticked for it here. Everything else is as for Venom: its Tasks and Incidents tools still run as **the person who
mentioned it** (it has no permissions of its own, and the team key's scope does not widen anything), a write waits for
their Confirm, and runs count toward the same daily limit. One message can mention several agents; each answers in its own
reply (at most three per message). Personal and Venom-daemon agents do not appear: they are programs that run elsewhere. A
disabled agent stops answering, and its reply says it is no longer set up here.

A service agent can also take a request on as work. With **Start work on a Venom Tasks issue** ticked, asking it to
build or fix something makes an issue from your message, delegates it to the agent, and shows the progress, the pull
request and the merge in the thread: [From a channel message to a merged pull request](/docs/cloud/chat-work.md).

## Run it on a computer

A channel owner can also tick **Use a computer** under the agent. The agent then works in its own browser on one of your
daemons, with logins that stay, instead of in the cloud: a mention starts a session next to that browser, and the
**Working…** reply lists what it did ("opened example.com", "read the page"). Before a risky action the reply asks you
(**Allow once**, **Allow on this site**, **Decline**); when it needs a login, a 2FA code or a CAPTCHA the reply says
**Needs you** and waits for **I'm done** or **Skip**. It uses its own daily limit (`daily_cap_usd` beside the tickbox,
else `chat_ai_daily_cap_usd`) and stops with a message when it is reached. The tools ticked above do not apply to a computer
run; the browser is its tool. Stop, re-run and editing work the same. See [Agents on computers](/docs/computers/agents-on-computers.md).
A channel without the tickbox runs in the cloud exactly as described on this page.

## Tools in other products run as you

A Tasks or Incidents tool runs through the product's own server with **your** role in the organisation, the same checks as
an API key of yours. It sees only what you may see and can change only what you may change; if the product refuses, the
reply says so in the product's words. The agent never has anything of its own there. Because the answer is posted in the
channel, whatever you may see in Tasks can end up in front of everyone in that channel: turn these tools on only in
channels where that is fine.

### Reading the other products

**Read from a Venom product** (`venom_read`) runs one read tool of another product, and **List a Venom product's read
tools** (`venom_tools`) tells the agent which ones there are and their arguments. Together they reach People, Funnel,
Support, Calendar, Flags, Incidents, Experiments, Computers, Notes, Tasks, Forms, the status page and the org's teams:
the same read tools an MCP client gets at `/workspace/mcp/<product>`. Ask "what's the latest on the Acme deal?" and the
agent can look the deal up in Funnel and the person in People.

They run through the [workspace MCP server](/docs/cloud/workspace-mcp.md) as you, the way an app you connected would with only the
`:read` scopes: your role decides what you see, a product your organisation does not have says so, and a tool that
changes anything is refused before it reaches the product. People and Calendar are yours alone, so a reply can show your
own contacts or meetings to the channel: tick these two tools only where that is fine. Chat (other channels), Comms
(messages and calls) and Admin (keys, budgets, the audit log) are left out on purpose.

Two tools rather than one per product tool keep the agent's prompt small: listing every read tool would add about 240
definitions (some 44,000 tokens) to each model call.

### Writes wait for you

When the agent wants to create an issue or comment, it does not do it. Its reply shows what it proposes, with **Confirm**
and **Decline** buttons that only the person who mentioned it can press. Confirm runs it then, as you, with your
permissions; the reply then says "Done as you" or why it did not run. A channel owner can tick **Run without asking the
person first** for a write tool; it then runs at once as the person who mentioned the agent. Proposals are never run
by anyone else, and a tool that is no longer allowed in the channel does not run even if it was proposed earlier.
Starting work is the one exception to running without asking: in a private channel it always waits for your Confirm,
because it copies your message out of the channel into Tasks.

## Stop, re-run and edit

While a reply says **Working…**, the person who asked, a channel owner or an org admin can press **Stop**. A finished,
failed or stopped reply has **Re-run**, which asks again for the same message and uses the daily limit like any run.
Editing your message that mentions @Venom runs it again for the new words, in the same reply.

## What it can and cannot see

- It reads as a member of that one channel. In a private channel it sees what the channel's members see; it never reads
  another channel.
- It only uses the tools the owner ticked. A model cannot widen that list or your permissions, whatever the conversation says; a write needs your Confirm unless the owner opted that tool out of asking.
- Only a person's mention starts it. Its own answers cannot mention anyone, so agents never trigger each other.
- Messages and notes are treated as untrusted text: it is told not to follow instructions inside them.

## Cost and limits

Each run goes through your organisation's own gear (the same routing as Support drafts), with at most four model calls,
and counts toward a daily limit: **Chat agents** on the [AI spend](/docs/cloud/ai-spend.md) page (`chat_ai_daily_cap_usd` in the
organisation settings), default $1, which owners and admins change there. When it is reached the thread says so. A run that takes more than five minutes is stopped and its reply says so. The cron `chat-agents` runs
anything still queued every minute.

## For agents and scripts

`chat_agent_get` and `chat_agent_set` (REST `GET`/`PUT /api/v1/chat/channels/{channel}/agent`; `tools`, and `auto` for the write tools that run unasked) read and change the setting;
`chat_agent_action_decide` (REST `POST /api/v1/chat/agent-actions/{action}`, `approve: true|false`) confirms or declines a proposed write
(the ids are in the reply's `meta.agent.pending`); `chat_agent_run` (REST `POST /api/v1/chat/messages/{message}/agent-run`,
`action: stop|rerun`) stops or re-runs a reply;
`chat_message_create` with a body containing `@[Venom](agent:venom)` starts a run, as the member the key acts for. A team
key (an agent) cannot start one.
