# Connect your AI client

Connect Claude, ChatGPT, Cursor or another AI assistant to Signal over Noise over MCP: the server address, its tools, how its proposals are approved, and calls to your team's own MCP servers.

Signal over Noise runs an MCP server, so the AI assistant you already use can work with your workspace. It works as you, in the workspaces you belong to. It can read, leave feedback, snooze cards and put questions to Ask directly; anything bigger is proposed, for someone who could make the change to approve in Signal over Noise.

## Connect

Add it as a custom connector in Claude, ChatGPT, Cursor, VS Code or Grok, then sign in with your Signal over Noise account. The server address is:

```
https://signalovernoise.app/mcp
```

Your assistant opens a Signal over Noise sign-in. One connection covers every workspace you belong to, and it can do only what your role in each workspace allows; see [Roles, approvals and access](/docs/access).

## What it can do

| Tool | What it does | Kind |
| --- | --- | --- |
| `ask` | Asks Signal over Noise a question. The conversation appears in your own Ask history. | Changes directly |
| `get_ask_answer` | Fetches an answer that was still being written. | Reads |
| `get_decision` | Reads one decision with the sources behind it. | Reads |
| `get_metric` | Lists the named metrics and queries them. See [Metrics and definitions](/docs/metrics). | Reads |
| `get_pulse` | Today's Pulse: the cards that need attention, with their sources. | Reads |
| `get_workspace_context` | Reads the company context pages. | Reads |
| `give_card_feedback` | Says whether a Pulse card was useful. | Changes directly |
| `list_connected_tools` | Lists the connected tools a call can be proposed to. | Reads |
| `list_decisions` | Lists recent decisions, pinned ones first. | Reads |
| `list_workspaces` | Lists the workspaces you belong to. Every other tool takes one. | Reads |
| `pin_conversation` | Pins or unpins one of your Ask conversations. | Changes directly |
| `pin_decision` | Pins or unpins a decision, so it is used as context first. | Changes directly |
| `propose_automation` | Proposes a new automation or a change to one. | Proposes for approval |
| `propose_automation_run` | Proposes running an automation now. | Proposes for approval |
| `propose_card_dismissal` | Proposes dismissing a Pulse card for the whole workspace. | Proposes for approval |
| `propose_card_restore` | Proposes bringing a dismissed card back. | Proposes for approval |
| `propose_context_page` | Proposes a new company context page or a new version of one. | Proposes for approval |
| `propose_integration_tool_call` | Proposes one call, with its exact arguments, to a connected tool that can write. | Proposes for approval |
| `propose_objective` | Proposes a new company objective or a change to one. | Proposes for approval |
| `propose_platform_drafts` | Proposes creating the ad drafts an Ask conversation prepared. Drafts are created paused. | Proposes for approval |
| `propose_recommendation_decision` | Proposes approving or ignoring a recommendation. | Proposes for approval |
| `rate_ask_answer` | Rates an Ask answer up or down. | Changes directly |
| `show_card_now` | Brings a snoozed card back now. | Changes directly |
| `snooze_card` | Hides a Pulse card for a few days. | Changes directly |

Members, integrations, settings and data erasure cannot be changed from an assistant at all.

## How a proposal is approved

A proposal changes nothing by itself. It returns a link to an approval page in Signal over Noise that shows the exact change. Someone who could make that change in the product approves or rejects it there, signed in as themselves, never through the assistant. Approving carries out exactly what was proposed, as the person who approved it.

## Your team's own MCP servers

A workspace admin can add an MCP server your team runs, in **Settings → Integrations** under **Your own tools**: its address, a name, what it is for, and the tools your assistants may use, at most 20. A tool left off that list is never called.

An assistant connected to the workspace lists the server's tools with `list_connected_tools` and proposes a call with `propose_integration_tool_call`. Nothing reaches the server until someone whose role can approve changes approves that call in Signal over Noise (see [Roles, approvals and access](/docs/access)), and an approved call is sent no more than once. Collaborators cannot see these servers or propose calls to them.

Only a server that answers without a key or a sign-in can be added. A key cannot be stored for your own server yet. The address must be the server's public `https://` address, with nothing after a `?` or `#` and no username or password in it. The server must also answer a single tool call on its own: each approved call is sent without first opening a session, so a server that needs one cannot be reached.

An admin can remove a server in the same place. AI clients are no longer offered its tools, and a call still waiting for approval will not run.

## Your data and your assistant

Members can connect AI assistants such as Claude, ChatGPT or Cursor. What an assistant reads goes to its provider under that provider's terms. Every question, proposal and change records the assistant that made it, by the name it gives itself. You can see and revoke your connected assistants in **Settings → Workspace profile**; a revoked one stays disconnected until you allow it again. A workspace admin can turn off access from AI clients for the whole workspace in **Settings → Privacy & consent**.
