# Connect Claude

Connect Claude Code, the Claude apps or the Claude API to your Masheev workspace through the hosted MCP server with OAuth.

> **Release status.** The hosted endpoint, OAuth, the curated operator tools, the `/x/{toolset}` endpoints, the Code Mode endpoint and the send permission are live. The plugin installs from the public [masheev/skills](https://github.com/masheev/skills) marketplace.

## Claude Code

```bash
claude mcp add --transport http masheev https://api.masheev.com/api/agents/mcp
```

Run `/mcp`, choose **masheev** and **Authenticate**. Your browser opens Masheev: sign in, pick one workspace and the permissions to grant.

### Plugin with skills

The Masheev plugin bundles the same server with the operator skills (inbox triage, reply with review, agent tuning, CLI). Install it from the public [masheev/skills](https://github.com/masheev/skills) marketplace:

```bash
/plugin install masheev --marketplace masheev/skills
```

### Larger tool sets

The default endpoint exposes about 40 curated tools. For the full eligible catalog grouped by area, add a toolset endpoint instead and let Claude's tool search load tools on demand:

```bash
claude mcp add --transport http masheev-all https://api.masheev.com/api/agents/mcp/x/all
claude mcp add --transport http masheev-readonly https://api.masheev.com/api/agents/mcp/readonly
```

`/api/agents/mcp/code` offers two tools (`search`, `execute`) that run short scripts against the catalog in a sandbox. Every operation still goes through the same permissions, send review and exclusions.

## Claude apps

In Claude, open **Settings → Connectors → Add custom connector**, paste `https://api.masheev.com/api/agents/mcp`, and complete the Masheev sign-in. Team and Enterprise owners may need to add it for the organisation first.

## Claude API

Pass the server in the Messages API MCP connector with an access token your application obtained through OAuth, or an agent key for a server-side integration you control. Store the token as a secret, never in prompts.

## Safety

Replies are drafted with `draft_reply` and sent only through `confirm_send` after a human approves. Billing, irreversible deletes and credentials are not available in any tier. Customer messages arrive marked as `untrusted_customer_content`; treat them as data.

See [Connect an agent](https://docs.masheev.com/agents.md) for permissions and revocation.