MCP

Real Bot: Put MCP Tools and Your Own Model Endpoints in the Hands of Persistent AI Teammates on macOS

For people who want persistent AI teammates running locally on macOS with their own OpenAI-compatible endpoints and MCP tools — but it's an unsigned alpha build that requires bypassing Gatekeeper on first launch, and new endpoints, operations outside the workspace, and outbound network requests all stop at an approval card awaiting your confirmation.

Official source · Blackman99/real-bot

Install

xattr -dr com.apple.quarantine "/Applications/Real Bot.app"

Who it's for

  • macOS users (builds exist for both Apple silicon and Intel) who want to run persistent AI teammates (bots) locally, bringing their own OpenAI-compatible model endpoints and MCP tools (stdio or Streamable HTTP).
  • People who are willing to accept alpha-stage instability and are comfortable handling unsigned-app first-launch issues.
  • Not a fit for: non-macOS platforms (no official builds for other systems are provided); anyone unable to accept bypassing Gatekeeper for an unsigned .dmg; or anyone seeking stability guarantees beyond the official documentation. The official evidence does not state pricing, telemetry details, or data-retention policies, so these cannot be confirmed.

First attempt in 3 minutes

The main path is to install the official alpha desktop app, then connect a model endpoint of your own:

  1. Prerequisites: a macOS device; any OpenAI-compatible endpoint plus its API key (the key is stored in the macOS Keychain and never enters chat).

  2. Install: download the latest alpha .dmg from GitHub Releases (separate Apple silicon and Intel builds) and drag it into /Applications.

  3. First launch: the build is unsigned, so Gatekeeper may block it. The officially suggested fix is right-click → Open, or run in a terminal:

    xattr -dr com.apple.quarantine "/Applications/Real Bot.app"

  4. First low-risk call: add your OpenAI-compatible endpoint following the official "Bring your own models" approach. Note that adding a new endpoint or MCP server is a gated action: an approval card appears for your confirmation — this is a deliberate confirmation point, and the change only takes effect after you approve.

  5. Success signal: the endpoint passes approval and enters the app, and the bot can use it for conversations.

Alternative path: the official repo also supports running from source (macOS, requiring Node 22+, pnpm 12.3.4, Bun 1.2+, Rust, and Xcode Command Line Tools, with cd real-bot && pnpm install && pnpm dev), but the dependency chain is longer and not ideal for a first attempt.

Use cases

Case 1: Connect your own model endpoint and get a conversation working. Scenario: you want your AI teammate to use an OpenAI-compatible service you already have rather than being tied to a single vendor. How: add the endpoint in the app (an approval card will appear; it only takes effect after you confirm), and per the official note the key goes only into the Keychain. Expected result: the bot can generate replies through your endpoint. The official docs do not list the exact field names in the configuration UI, so go by what the app actually displays.

Case 2: Attach an MCP server for all bots to use. Scenario: you already have an MCP server exposed via stdio or Streamable HTTP and want bots to call its tools. How: add the MCP server in the app — this is likewise a change that requires approval, and you allow it when the confirmation card appears. Expected result: per the official docs, an MCP server is "available to every bot". How the tool list within a single MCP server is displayed is not detailed in the official README; go by what the app shows.

Boundaries of use

  • Writes and side effects: this is an agent app that can take actions. Per the official docs, adding a new endpoint or MCP server, operations outside the workspace, and outbound network requests all stop at an approval card awaiting your confirmation. There is no further explanation of approval granularity beyond the official wording.
  • Local state writes: API keys go into the macOS Keychain; bots persist within the workspace ("persistent AI teammates"), but the specific storage location for this persistence is not stated officially.
  • Stage limitations: the project is in Alpha; the .dmg downloaded from Releases is unsigned, and bypassing Gatekeeper means forgoing macOS signature verification for this app — the risk is yours to bear.
  • Not yet confirmed: platform support (official docs mention macOS only), pricing, privacy and data handling, Windows/Linux support, and the exact fields for MCP server configuration — none of these are covered in the evidence, so verify them yourself in the app or the repo before use.