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 ~/.agentty/credentials/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, DeepSeek, Kimi, Gemini, 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, DeepSeek, Kimi, Google Gemini, xAI Grok, Mistral, Groq, OpenRouter, Together, Cerebras, Fireworks, 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, DeepSeek, Kimi, Google Gemini, xAI Grok, Mistral, Groq, OpenRouter, Together, Cerebras, Fireworks, 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.

Can I use Kimi or DeepSeek?

Yes — both are first-class. Kimi: agentty login → Kimi signs you in with your Kimi plan over OAuth (no API key), then agentty --provider kimi. DeepSeek: set DEEPSEEK_API_KEY and run agentty --provider deepseek -m deepseek-v4-pro (or deepseek-reasoner for the thinking model — its reasoning streams live). 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 a sandbox the kernel enforces, so it holds even if the command tries to escape. 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 files under ~/.agentty/threads/: one append-only .jsonl log per thread plus small index sidecars. Safe to inspect, back up, or delete. Move them with AGENTTY_THREADS_DIR, expire old ones with AGENTTY_THREADS_KEEP_DAYS. agentty config shows every folder agentty uses and its size; Storage has the full map.

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.