Agneta

Give Mac apps an AI agent you control.

Agneta is an open-source menu bar app for macOS. Connect the AI you already pay for once: a ChatGPT or Claude plan, or an OpenAI or Anthropic API key. Then decide which apps may use it. Agneta runs each task on your Mac, in a sandbox, on your account, and shows you every task with a stop button.

Agneta’s menu bar panel: two running tasks from Example App, each with a stop button, and three connected providers with the engine each runs on Agneta’s menu bar panel: two running tasks from Example App, each with a stop button, and three connected providers with the engine each runs on
The menu bar panel: each running task, its app and provider, and a stop button.
An app asks to use Agneta. The window shows what it may do and that it’s verified, and the user ticks which providers it may use An app asks to use Agneta. The window shows what it may do and that it’s verified, and the user ticks which providers it may use
An app asks first. You see what it may do, and tick the plans and keys it may use.

For you, with your own AI plan

Use the plan or key you already pay for in the apps you choose, see what each one does with it, and stop it.

  • An app has to ask first. You see the tools and folders it wants and whether its code signature checks out, then allow it once or always.
  • You choose which of your plans and API keys each app may use, and can remove its access at any time.
  • Every task shows the app it belongs to, the provider it runs on and what it costs.

For app developers

Agent features without running an agent loop, building sign-in for every provider, or putting a model bill on your own credit card.

  • Speak the Agent Client Protocol over a local WebSocket. No SDK is needed, and a sandboxed Mac App Store app can do it too.
  • Keep your tools and data in your app, served as an MCP server on loopback. The agent uses them on the user’s account.
  • Send a policy with each session: tools, folders, network, approval and limits. Agneta enforces it with sandboxes, or refuses it.
  • Usage comes back with every result, and every task leaves a full trace, live or later.

Read the developer guide

How it works

Your appits own tools, as an MCP server
ACP over ws://127.0.0.1
Agnetaaccounts · consent · policy · limits · traces
ACP over stdio
Agent engineGoose, OpenCode, Claude Code or Codex, sandboxed
The engine calls your app’s tools over MCP, on loopback
Your app talks to Agneta in ACP over a local WebSocket. Agneta talks to a sandboxed agent engine in ACP over stdio. The engine calls your app’s own tools over MCP on loopback.
  1. Pair once. The app connects and calls agneta/pair with the scope it needs (tools and folders). Agneta identifies the process by its code signature; the user sees the scope, ticks the providers it may use, and allows it once or always. The app keeps a token in its Keychain, and agneta/access tells it which providers and engines it may use.
  2. Start a session. It calls session/new with a provider and, if it wants, an engine, a workspace, the app’s MCP server and a policy, for example {"tools": "workspace", "network": "none", "limits": {"maxTurns": 200}}. Agneta refuses any policy it can’t enforce rather than quietly weakening it.
  3. Prompt. Agneta queues the turn within the account’s limits and starts or reuses an engine under the right sandbox profile. It streams the engine’s ACP updates back, answers permission requests from the policy (or asks the app), and returns usage with the result.
  4. Watch and stop. The task appears in the menu bar panel, its trace is written as it runs, and the user can stop it at any time.

What stays on your Mac

Agneta has no servers of its own and sends no telemetry. What leaves your Mac is the task itself: its prompts, and what its tools return, go to your AI provider over HTTPS, because that’s how the model sees it.

Your sign-ins
Plans are signed in by your own Codex or Claude Code, which keep and refresh their credentials. Agneta never holds a plan’s tokens, and API keys stay in your Keychain.
Loopback only
Agneta listens on 127.0.0.1 and refuses requests from web pages. Nothing listens beyond your Mac.
Sandboxed engines
Each engine runs under a macOS sandbox profile: the task’s workspace, the folders its policy names and what the engine itself needs, and nothing else in your home folder.
Commands, stricter still
Shell commands run confined to the workspace, with no network when the policy says none.
Traces you keep
Every task leaves a trace of its prompts, tool calls, model calls and decisions, readable only by you and kept 30 days by default.
Updates, once you allow them
Release builds ask before they first check GitHub for updates. The check carries Agneta’s version and nothing of yours.

Privacy and threat modelHow engines are confined

Providers and engines

Agneta doesn’t run the agent loop itself. It hosts proven open-source and vendor agents behind one boundary, and each app picks the engine for its tasks among those you allow.

ProviderEnginesSign-in
ChatGPT planCodexYour installed Codex’s own login, kept per account
Claude planClaude Code, GooseYour installed Claude Code’s own login, kept per account; both engines run that Claude Code
OpenAI API keyGoose, OpenCode, CodexKey in the Keychain
Anthropic API keyGoose, OpenCode, Claude CodeKey in the Keychain

Without Claude Code or Codex, Agneta still works with an OpenAI or Anthropic API key, on Goose (bundled) or OpenCode (downloaded when first needed). A Claude plan needs Claude Code (the Claude desktop app includes it), and a ChatGPT plan needs Codex 0.166 or later.

What each engine can do

What it is not

Install

Version 0.3.0, 18 November 2026. Signed with a Developer ID and notarized by Apple. Needs macOS 14 or later.

Not sure which you have? Choose Apple menu → About This Mac: Chip: Apple M… means Apple silicon, Processor: … Intel means Intel.

Then

  1. Open the DMG and drag Agneta into Applications.
  2. Open Agneta from Applications. It lives in the menu bar (the linked rings beside the clock), not in the Dock.
  3. Click the menu bar icon and connect a provider: a ChatGPT plan (needs Codex 0.166 or later), a Claude plan (needs Claude Code; the Claude desktop app includes it), or an OpenAI or Anthropic API key.
  4. To see what an app does with it, choose Run a Test Task from the panel’s ⋯ menu.

Or with Homebrew

brew install olofd/tap/agneta

Updates

Agneta updates itself with Sparkle. It asks before it first checks GitHub, then checks about once a day and installs an update only when no task is running; Settings → General turns checks or installs off. Homebrew leaves updating to it.

Check a download

Each release lists every file’s SHA-256 in SHA256SUMS.txt, and GitHub records how each was built: gh attestation verify checks a DMG against this repository.

Or build it

With Xcode 26 or later, from the source: Build and run.

Status

Status: early. Agneta runs real tasks on real ChatGPT and Claude plans. Before you rely on it, know that:

  • the app protocol may still change before 1.0;
  • an engine can reach any HTTPS host, not only its provider’s (a per-provider host allowlist isn’t built yet);
  • the tests, the sandbox acceptance test among them, pass on macOS 15 and 26, Apple silicon and Intel (CI.md), but the app has been used day to day only on an Intel Mac with macOS 26, and macOS 14 is untested.

Add it to your app

Your app is an ACP client and Agneta is the ACP agent. Connect to ws://127.0.0.1:47321/acp, pair once, and start sessions with your own tools and a policy. Any WebSocket and JSON-RPC 2.0 code will do.

Pairing, once

→ {"jsonrpc": "2.0", "id": 1, "method": "agneta/pair",
   "params": {"name": "Example App", "scope": {"tools": "workspace", "folders": ["/Users/me/Documents/Example Project"]}}}
← {"jsonrpc": "2.0", "id": 1, "result": {"token": "…", "remembered": true, "access": {…}}}
Project Chat: a conversation with the agent, the project tools it called, and a proposed note with Dismiss and Save note buttons
Project Chat, the example app: it pairs with Agneta, streams a conversation, and lets the agent propose a note for you to review.