FAQ

Quick answers to the questions people ask most.

Does it work with my Claude Pro/Max subscription?

Yes — OAuth against your existing Pro/Max plan is supported (no extra billing, same account). But it's not the only path or even the recommended one: agentty is bring-your-own-model, so most people run it with an API key or a local Ollama model. See Authentication.

Will using my Claude subscription with agentty get my account banned?

The honest answer: agentty is a third-party client, so it carries the same footing as any third-party client that speaks to Anthropic's API. What agentty actually does is deliberately unexotic — it completes the same OAuth flow and uses the same CLAUDE_CODE_OAUTH_TOKEN mechanism Claude Code itself uses, sends ordinary Messages-API requests over HTTPS, and adds nothing that spoofs, scrapes, or circumvents rate limits. Credentials live only in ~/.config/agentty/credentials.json (mode 0600); nothing is sent anywhere except Anthropic.

We're not affiliated with Anthropic and can't speak for their enforcement, and their terms can change. If you want zero ambiguity, use an ANTHROPIC_API_KEY (pay-as-you-go, unquestionably in-bounds) or point agentty at another provider — OpenAI, Groq, OpenRouter, or a local Ollama model — with --provider. See Authentication.

Is it really a drop-in for claude-code?

It targets the same workflow — a coding agent in your terminal — as a single native binary, and it can use the same Claude auth. But it's bring-your-own-model: alongside Claude it runs against OpenAI, Groq, OpenRouter, Together, Cerebras, and local Ollama models. See the full agentty vs Claude Code comparison and Providers & Models.

Can I use models other than Claude?

Yes. Pass --provider to point agentty at OpenAI, Groq, OpenRouter, Together, Cerebras, a local Ollama model, or any raw host:port. It persists like -m, and you can switch backends live in-app with ^P. See Providers & Models.

Do I need Node or Python?

No. agentty is a single static C++26 binary. No Node runtime, no npm install, no Python, no Electron.

What platforms are supported?

Linux, macOS, and Windows — all built and tested daily. Prebuilt binaries ship for Linux (x8664, aarch64) and Windows (x8664); macOS builds from source in seconds. It also builds natively on Termux/Android (no root, no proot) — see Installation.

How is it sandboxed?

Every shell/build call runs in bwrap (Linux) or sandbox-exec (macOS). The workspace is read-write, system libs read-only, and ~/.ssh / /etc / other projects are blocked. Windows and unrooted Android/Termux run unsandboxed (no user-namespace backend available there); agentty detects this and prints sandbox: unavailable, running unsandboxed.

Can I run it on a machine with no internet?

Yes. agentty airgap user@host relays traffic from your laptop over SOCKS5-over-SSH, with TLS pinned end-to-end. See the air-gap guide.

Can I use it inside my editor?

Yes. agentty acp runs agentty as an Agent Client Protocol agent inside Zed — streaming responses, inline diffs, native permission prompts, and session reload. Any ACP client works; it's the same engine as the TUI, just over JSON-RPC on stdio.

Where are my conversations stored?

As plain JSON, one file per thread, under ~/.agentty/threads/. Safe to inspect, back up, or delete.

What happens when a conversation runs out of context?

Two options. agentty auto-compacts in the background as a thread nears the window — it summarizes the older turns (on the cheapest capable model) and keeps going in the same thread. Or you can fork (Ctrl+K → Fork thread): that branches into a fresh thread carrying near-zero context, writing the parent's full transcript to disk for the model to read on demand. Compaction keeps one thread alive by compressing its history (lossy); fork starts clean and leaves the history verbatim on disk. Fork when you want a clean slate without losing the ability to look back.

Is it stable / production ready?

It's pre-1.0 and moving fast, but the core loop, tools, streaming, auth, and persistence all work and get daily smoke testing on Linux. Treat it as a capable beta.

What license is it under?

MIT. Use it, fork it, ship it. See the license page.

How do I fix or improve these docs?

Every page here is a Markdown file in the agentty repo under docs/website/. Click Edit this page at the bottom of any doc, or open a PR against that file — the site rebuilds itself from the Markdown automatically, so a merged change goes live on its own.