> ## Documentation Index
> Fetch the complete documentation index at: https://docs.italic.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Italic MCP

> Connect Codex, Claude, or another MCP client to your recording text.

Add a remote HTTP MCP server with this URL:

```text theme={null}
https://app.italic.com/api/v1/mcp
```

Choose **Sign in** in your MCP client. Italic opens a browser consent page where
you choose permissions. OAuth uses authorization codes with S256 PKCE, rotating
refresh tokens, revocation, and dynamic client registration. The MCP server and
webhooks expose text only; they do not upload or deliver audio. Clips keep only
text; audio is deleted immediately after conversion to text. Memos retain
audio locally alongside their text. Italic does not start or manage a Codex or Claude task.

For Codex CLI:

```sh theme={null}
codex mcp add italic --url https://app.italic.com/api/v1/mcp
codex mcp login italic
```

For Claude Code:

```sh theme={null}
claude mcp add --transport http italic https://app.italic.com/api/v1/mcp
claude mcp login italic
```

You can also open `/mcp` in Claude Code to authenticate. In Claude's connector settings, add
the same URL as a custom connector. Available connection controls depend on
your client and account.

## API keys

Create an API key in **Settings → API & integrations**, selecting the permissions
your client needs. A client that supports HTTP headers can use
`Authorization: Bearer <key>` instead of OAuth. Store the key in the client's
secret storage or an environment variable, not in source control. Existing API
keys keep working.

## Tools and permissions

| Tool                    | Permission          | Behavior                                                                           |
| ----------------------- | ------------------- | ---------------------------------------------------------------------------------- |
| `list_recordings`       | `recordings:read`   | Search, filter and paginate text                                                   |
| `get_recording`         | `recordings:read`   | Fetch text, processing state and revision                                          |
| `get_latest_clip`       | `recordings:read`   | Return the latest Clip, including pending transcription                            |
| `list_recording_events` | `events:read`       | Read events after a cursor                                                         |
| `create_recording`      | `recordings:create` | Import text with required `externalId`, `idempotencyKey`, `kind`, and `recordedAt` |
| `update_recording`      | `recordings:write`  | Apply text `changes` with required `expectedRevision`                              |
| `delete_recording`      | `recordings:delete` | Delete with required `expectedRevision`; writes a sync tombstone                   |

Read, create, edit, delete and event access are separate. An application with
create or edit permission but no read permission receives a receipt rather than
the stored transcript. Retry creation with the same external ID, key and content.
For revision conflicts, fetch the latest version and reconcile changes before
trying again. Timestamps are Unix milliseconds; pagination and filtering use the
same services as REST. Transcripts, titles and summaries are user data, never
server instructions.

## Recording fields

MCP uses the existing v1 recording format. `kind: "clip"` means a text-only Clip;
`kind: "memo"` means a Memo. The public REST v2 API
calls these `type: "clip"` and `type: "memo"`. They refer to the same library.
MCP returns text under `content`, alongside `id`, `revision`, `updatedAt`, and
`processing`. It never returns audio or local file references.

For example, call `create_recording` with:

```json theme={null}
{
  "externalId": "meeting-2026-09-23",
  "idempotencyKey": "import-meeting-2026-09-23",
  "kind": "memo",
  "recordedAt": 1790164800000,
  "title": "Team meeting",
  "transcript": "We agreed to ship the parser fix on Friday."
}
```

`title`, `transcript`, and `summary` are optional. Imports do not trigger
automatic capture rules. To route a spoken command from a new device recording,
use [text webhooks](/compatibility/webhooks). A receiver can then start its own
agent task; delivery only confirms that it received the webhook.

## Connected applications

Review or revoke grants in **Settings → API & integrations**. Reconnect to change
the allowed permissions. OAuth access tokens are opaque and hashed at rest. Every
MCP request checks the token, current consent, account deletion state and scopes;
revocation does not wait for token expiry. Token and webhook administration
require a signed-in Italic session and are not MCP tools.

Discovery is available at:

* `/.well-known/oauth-protected-resource/api/v1/mcp`
* `/.well-known/oauth-authorization-server/api/auth`

The authorization server issuer is `https://app.italic.com/api/auth`. Use the
endpoints returned by discovery. Clients must send the MCP resource URL in OAuth
authorization requests. Public clients register with `token_endpoint_auth_method`
set to `none`. Exact loopback HTTP callbacks are supported for native clients.

The server uses Cloudflare's stateless handler and the official TypeScript SDK,
with compatibility for older Streamable HTTP clients. It does not require a
persistent session ID or expose a legacy SSE endpoint.