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

# Claude Code

> Connect Claude Code to a BeeOS agent with MCP and add reusable workflow guidance with Skills.

Claude Code connects to BeeOS over **MCP**. A Claude Code Skill is optional: it
teaches Claude when and how to use the BeeOS tool, but it does not carry network
traffic or replace MCP authentication.

## Add the MCP server

Use the BeeOS agent ID in the remote Streamable HTTP endpoint:

```bash theme={null}
claude mcp add --transport http beeos-device \
  "https://mcp.beeos.ai/${BEEOS_AGENT_ID}/mcp"
```

Run `/mcp` inside Claude Code and complete the browser OAuth flow. For CI or a
non-interactive environment, use a scoped API key header instead:

```bash theme={null}
claude mcp add --transport http beeos-device \
  "https://mcp.beeos.ai/${BEEOS_AGENT_ID}/mcp" \
  --header "Authorization: Bearer ${BEEOS_API_KEY}"
```

<Warning>
  Shell expansion can place a literal key in Claude Code's configuration. Prefer
  OAuth for developer workstations and use your secret manager's injection
  mechanism for automation.
</Warning>

Verify the connection:

```bash theme={null}
claude mcp get beeos-device
claude mcp list
```

Then ask Claude to perform a low-risk action, for example: `Use the BeeOS device
tool to report the battery status.` Claude will discover the agent through
`tools/list` and invoke it through `tools/call`.

## Add a Skill

A project Skill can encode your safety and verification policy:

```markdown theme={null}
---
name: operate-beeos-device
description: Operate the connected BeeOS device for explicit user requests.
---

Use the BeeOS MCP tool only when the user explicitly requests a phone action.
Before purchases, sending messages, deleting data, or changing an account,
describe the pending action and ask for confirmation. After an action, verify
the visible result before reporting success.
```

The Skill controls model behavior; the MCP server remains the source of tool
schema, execution, and results. Do not put credentials in `SKILL.md`.

## Team configuration

Use project scope only when every contributor should see the server definition:

```bash theme={null}
claude mcp add --scope project --transport http beeos-device \
  "https://mcp.beeos.ai/${BEEOS_AGENT_ID}/mcp"
```

Claude Code writes project-scoped MCP configuration to `.mcp.json` and asks
each user to approve it. Commit the endpoint only; each user should authenticate
with their own BeeOS identity.

See [MCP authentication](/mcp/oauth) and the
[Claude Code MCP documentation](https://docs.anthropic.com/en/docs/claude-code/mcp).