Agneta

Privacy and threat model

What leaves your Mac when an app runs a task through Agneta, what stays on it and for how long, what each button removes, and what Agneta defends against and what it doesn’t.

What leaves your Mac

WhatWhere it goesWhy
The task’s prompts and instructions, the conversation so far, and what its tools returned (including the files an engine read)Your AI provider (OpenAI or Anthropic), through the engine, over HTTPSThat is how the model sees the task
Tool calls to the app’s own toolsThe app, on this Mac (loopback)The app’s tools run in the app
A request for the public model catalog, with nothing of yours in itmodels.dev, from Goose when it startsGoose looks up model names and limits there and has no setting to stop it (Engines)
Update checks, once you allow them: a request for the update feed, which carries Agneta’s version and nothing of yours; and the update itself, when one installsGitHub (github.com and its download host), from release builds onlyAgneta asks before its first check and checks about once a day; Settings → General turns checks off. Sparkle’s system profile, which would send the Mac’s model and macOS version, stays off (RELEASING.md)
Downloads of components you choose: Claude Code’s adapter when you add a Claude plan, OpenCode when it becomes a provider’s default or a task asks for it, Codex’s runtime (Node and codex-acp) when Codex is first usedGitHub (github.com and its download host)Each is a pinned release file, checked against its SHA-256 before use; the request carries nothing of yours
Nothing else from AgnetaAgneta has no servers and sends no telemetry

Engines are run with their own telemetry, update checks, analytics and feedback turned off where they allow it (Engines). On Codex, web search (which runs on OpenAI’s servers) and the features that bring tools of their own (the ChatGPT account’s connectors, plugins, browser and computer use) are off. What your provider keeps, and for how long, is set by your account with them.

What stays on your Mac

WhatWhereKept
Task history: titles, apps, accounts, turns, usage, outcomeAgneta’s databaseAs long as task history (30 days by default; Settings → Privacy)
Each task’s trace: prompts, tool calls and results, model calls, decisions, timingTraces/, one owner-only file per taskAs long as its history
An engine’s own record of the conversation, so the task can resumeThe engine’s home (Engines/), Claude Code’s transcripts in the account’s folder, Codex’s sessions in the account’s CODEX_HOMEAs long as its history: each goes with its task. A task that ran on Goose or OpenCode without built-in tools shares its account’s engine: it’s a row in that engine’s own database, which goes when the account is signed out
Agneta’s own log: events, ids, names, counts and timing, never content, tokens or keys~/Library/Logs/Agneta/7 days
Downloaded components (Claude Code’s adapter, OpenCode, Codex’s runtime)Runtimes/ in Agneta’s data folderUntil Reset; an update keeps them while their pin is unchanged
Update settings, the last check’s time, and an update downloaded but not yet installedAgneta’s preferences; ~/Library/Caches/com.objectivecoder.agneta/Until the update installs, or Reset
API keys and the diagnostics tokenAgneta’s Keychain itemsUntil removed
Plan sign-insCodex’s and Claude Code’s own entries in the login Keychain, one per accountUntil signed out (Sign Out and Reset run each one’s logout)
What tasks wroteThe workspace the app named, or the app’s own workspace (… Workspaces/<grant>)Until you delete it, or Reset for the app’s own workspace

Everything Agneta writes is readable only by your user account (folders 0700, files 0600).

What each action removes

ActionEffect
Stop (a task)The task and its engine stop now. Its history and trace stay.
Pause (an account)Nothing new starts on the account; a running turn finishes. Nothing is removed.
Remove Access (an app)The app’s token stops working, its tasks stop and its connections close; the grant is deleted. Its tasks’ history and its own workspace stay.
Sign Out (an account)Its key or sign-in is deleted (Claude Code signs out of its own entry), and so are the account’s folder and its engines’ homes. Its tasks’ history and traces stay until they expire.
Clear HistoryEvery finished task’s history row, trace and engine home is deleted, and so are Claude Code’s transcripts and Codex’s session files of those tasks.
Reset AgnetaEverything above, and every account, grant, setting, log, cache, the apps’ workspaces, Agneta’s Keychain items and its login item.

The diagnostics token

Off until you create it (Settings → Advanced → Diagnostics). A tool holding it, such as Claude Code, can read every task’s trace on this Mac, prompts and tool results included, through a local MCP server on loopback that refuses browsers. It can’t start tasks, change settings or reach accounts. Turning diagnostics off, or making a new token, stops the old one at once.

Sharing a trace

agneta-diag export <task> --out <file> --redact-content (or Export Diagnostics in a task’s details, with “Leave out content, names and paths” on) writes a copy fit to attach to a bug report. It keeps the trace’s structure, timings, sizes, token counts and costs, and what’s structural: record kinds, statuses, decisions, model and engine names. Everything else is removed and replaced with its size: prompts, tool input and output, instructions, error text, app and account names, identifiers and every path. Tool and server names become tokens, the same within one export and different in the next. Without --redact-content the export is the whole trace: share it only with someone you’d show the task to.

Threat model

What Agneta protects: your AI accounts (credentials and spend), your files, and what your tasks contain.

WhoWhat Agneta doesStatus
An app you allowedIt gets the accounts you picked, at most the tools and folders it asked for and you saw, and a token. Every session is held to that grant; anything beyond it is refused. You see and can stop each task. A verified app’s token works only for that app’s signatureDefended
A program you didn’t allow (another app, a sandboxed app, a script)It can’t use Agneta without pairing, and pairing needs your approval in Agneta’s own window. An app Agneta can’t verify must show a code you compareDefended
The model, or text it reads (a prompt injection in a file, a web page an app fetched, a tool result)The engine and its commands run sandboxed: only the policy’s folders, no network for commands while the policy says none, nothing launched outside. Built-in tools beyond the policy are refusedDefended, with the limits below
A web page in your browserAgneta’s ports refuse requests carrying an Origin, which browsers always sendDefended
Other users of this MacEverything Agneta stores is owner-only; it listens on loopback onlyDefended
Someone on your networkNothing listens beyond loopback; engines reach providers over TLSDefended
Unsandboxed code running as youIt can already read your files and drive your apps. Pairing slows it down (it must get past the approval window) but can’t stop itOut of scope
Your AI providerIt sees what the model is sent, under your agreement with itOut of scope

Limits

  • An engine can reach any host on port 443. macOS’s sandbox filters by port, not host, and engines need HTTPS for their provider. A tool can’t use that, but the engine process itself could (Containment).
  • The app’s own tools run in the app, with the app’s permissions: Agneta doesn’t confine them.
  • Claude Code and Codex run their own tools inside the engine’s sandbox, checked by Agneta from their permission requests rather than by the stricter command sandbox.
  • Development builds are signed ad hoc, so an app built that way can’t be verified; it pairs with a code.
  • The sandbox relies on sandbox-exec, which macOS marks deprecated (Containment).

This page is docs/PRIVACY.md in the repository, on main. History