Apps (MCP)

Apps connect Agentbits to MCP (Model Context Protocol) servers, the same kind of connectors you may know from Claude Desktop: mail, calendar, design tools, Notion, databases and more. Agents never talk to the server directly. Every call goes through Agentbits, which checks the agent’s access and the tool’s policy, asks you when needed, and writes the call to the audit log.

Add an app

Open Organization → Apps and click Add app. You have four options:

Option	Use it for
Local program (stdio)	A server that runs on your computer, started by a command such as npx or uvx.
Remote server (URL)	A server reachable over HTTP at an address.
Paste JSON config	A config snippet copied from a README or another app.
Import from Claude Desktop	The servers you already set up in Claude Desktop.

Local program

Fill in Name, What is it for? (agents see this), Command (for example npx), Arguments (one per line), an optional Working folder and Environment variables for API keys and tokens.

Warning: Agentbits runs this command on your computer with your user’s rights and can’t restrict what it does. Only add servers you trust, such as official ones or ones from a source you know. Many need Node.js (npx) or uv (uvx) installed; Agentbits doesn’t install them for you.

Remote server

Fill in Name, the description, the Server address (for example https://example.com/mcp) and, if needed, Extra HTTP headers such as Authorization: Bearer …. Agentbits uses Streamable HTTP and falls back to the older SSE transport when the server needs it.

For servers that ask you to log in (most hosted apps), turn on Sign in with OAuth in the browser, save, then click Sign in. Your browser opens the app’s sign-in page, and Agentbits stores the token encrypted on this computer. Only sign in to servers you trust, and check that the page belongs to the service you expect before you enter a password.

Agents never open a browser. If a token is missing or can’t be refreshed, the app shows sign-in needed and the agent gets an error until you sign in again.

Paste JSON or import from Claude Desktop

Paste JSON config reads the common formats, for example {"mcpServers": {"name": {"command": "npx", "args": […]}}}, a servers block, or a single "name": {…} entry copied from a README. Click Read, review the servers found and add them.

Import from Claude Desktop reads Claude Desktop’s claude_desktop_config.json:

• Windows: %APPDATA%\Claude\
• macOS: ~/Library/Application Support/Claude/
• Linux: ~/.config/Claude/

An app whose name already exists is skipped (already added (skipped)). New apps are tested in the background.

Secrets

Values of environment variables and headers, and OAuth tokens, are stored encrypted with your operating system’s secure storage. They are never shown again; leaving a value empty when you edit keeps the old one.

Tools and policies

Click Connect & list tools to connect and load the server’s tools. The status next to the app shows connected, error, not tested, off or sign-in needed.

On the Tools tab, each tool has a Policy:

Policy	What it does
Runs automatically	Agents with access call it without asking.
Ask me first	Every call waits for your approval.
Off	Agents don’t see the tool at all.

By default, tools the server marks as read-only (and not destructive) run automatically, and everything else is set to Ask me first. These hints come from the server itself, not from Agentbits, so treat them as claims. If you’re not sure about a tool, keep it on Ask me first.

Who can use an app

Nobody has access by default. Grant access on the app’s Access tab (Allow all and Allow none are shortcuts), or from the agent’s page under Organization → Agents in the Apps section. An agent with access can use all of the app’s tools that aren’t Off, each under its policy.

The In use switch turns the whole app off: when off, no agent can use it.

Approvals

When an agent calls an Ask me first tool, the request appears in the approval panel at the bottom left (titled Approval required) with the tool’s arguments, plus a toast. Approve it or reject it. If you don’t answer within 10 minutes, it is rejected. If the agent loses access, or the app or agent is changed or deleted while the request waits, it is rejected. After you approve, Agentbits checks again that the agent still has access and the tool isn’t Off before the call runs.

Audit log

The app’s Audit log tab records every agent call (the tool, the result and how long it took) and your own changes: app added, changed or deleted, tool policy changes and access changes. Deleting an app keeps its audit entries.

Providers

Provider	App support
Anthropic API, OpenAI API	Yes, as native tool calls.
Claude Code CLI	Yes, through Agentbits’ own MCP bridge.
Codex CLI	Yes, through the same bridge, for agents that have an app.
Demo	Tries a tool of an app named in the task, and asks for approval when the task asks for a change.

Agentbits ignores a project’s own .mcp.json when it runs Claude Code, so a repository can’t hand an agent its own servers. App support for Claude Code and Codex CLI has so far been tested only with simulated CLIs, not the real ones. See Providers.

Limits

• Only MCP tools are supported. Resources, prompts, sampling and elicitation are planned.
• OAuth sign-in has been tested only against a reference test server, not yet with real services such as Notion or Linear. Services that require a pre-registered client ID are not supported.
• App results reach the model marked as untrusted data, but that is not a guarantee against prompt injection: an email body, for example, can contain instructions. Keep tools that change things on Ask me first.

See also Permissions and Security.