> ## Documentation Index
> Fetch the complete documentation index at: https://speaker-weave.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Chat agent

> An Every-style in-app agent — streaming, @-mentions, entity badges, and approval gates. Bring your own OpenAI or Anthropic key.

The chat agent is a right-side panel on every organizer page: threads, streaming responses, `@`-mention any submission/speaker/session as context, clickable entity badges that navigate the app, and inline Approve/Deny confirmation before anything sensitive happens (emails, decisions, publishing, deletes).

## One switch

The agent is **dormant by default** and activates only when a provider key exists:

| State                | Configuration                                                                           |
| -------------------- | --------------------------------------------------------------------------------------- |
| Off (default)        | No key set — no UI renders, no routes mount, no AI SDK imports                          |
| OpenAI (recommended) | `OPENAI_API_KEY` — OpenAI Agents SDK, `gpt-5.6-luna` at `xhigh` reasoning               |
| Anthropic            | `ANTHROPIC_API_KEY` (or `ASSISTANT_PROVIDER=anthropic`) — same event protocol on Claude |
| Force off            | `ASSISTANT_ENABLED=false` even with keys present                                        |

The frontend follows automatically via `GET /api/agent/capabilities` — there is no second flag.

## Architecture

`api/agent/service.py:run_turn` owns the complete provider-neutral turn lifecycle: thread claims and persistence, history, prompts, runtime events, tool dispatch, permission cleanup, and the final reply. In-app Ask is an SSE adapter over that function. The [Slack agent](/ai/slack) is an Events API adapter over the same function, so both transports use the same tools, MCP connectors, thread store, and permission gate; Slack approval requests arrive as Approve/Deny buttons and Slack threads appear in Ask history.

The provider harness is shared — tool registry, event vocabulary, permission gate, thread persistence — with the model loop as the only swappable part. The full provider-swap notes (including what differs when running on Claude) live in the repo at `docs/chat-agent.md`.

## MCP connectors

Organizations can attach external tool servers from Settings → Integrations: [Every](https://every.ai) ships as the first preset, and any Streamable-HTTP MCP server can be added with OAuth (discovery + dynamic client registration + PKCE), a static bearer token, or no auth. External tools are namespaced `mcp__<connector>__<tool>` and every external mutation passes the same Approve/Deny gate. See [MCP connectors](/integrations/mcp-connectors).