Kimi K2 and DeepSeek in agentty — sign in with Kimi, key in DeepSeek, code from your terminal

If you run Kimi or DeepSeek, you now have a fast, native coding agent built for them. As of agentty 0.4.0, both are built-in providers — not a proxy, not a wrapper, not a plugin. Kimi you sign in to with your existing Kimi plan over OAuth (no API key, just like "Sign in with ChatGPT"). DeepSeek takes a DEEPSEEK_API_KEY. Either way you get the whole agent — streaming output, native tool-calling, file edits, shell, retrieval, and live-streamed reasoning — from a single static binary with a millisecond cold start. No Node, no Python, no Electron, no npm install.

Sign in with Kimi (no API key)

Have a Kimi plan? Use it directly. agentty implements Kimi Code's OAuth device flow (RFC 8628) against auth.kimi.com and talks to Kimi's OpenAI-compatible coding API — so you run Kimi K2 with no key to manage:

# 1) install agentty (single static binary)
curl -fsSL https://agentty.org/install.sh | sh

# 2) sign in — pick Kimi at the menu; it prints a one-time code + URL
agentty login

# 3) run agentty on Kimi
agentty --provider kimi

agentty login shows a one-time code and opens the Kimi authorization page. Over SSH? Press c to copy the code (OSC 52 → your local clipboard), paste it in any browser, approve, and agentty switches the moment it's authorized. The token is stored encrypted and refreshed automatically mid-session — and you can hold multiple Kimi accounts and switch between them live in-app with ^P.

That's the whole setup: no MOONSHOT_API_KEY, no base-URL juggling. kimi is a real provider preset, so --provider kimi persists between runs — after the first launch you just type agentty.

Bring a key for DeepSeek

DeepSeek is a first-class preset too. Set your key and point agentty at it:

# general coding — DeepSeek V4
export DEEPSEEK_API_KEY=sk-deepseek-YOURKEY
agentty --provider deepseek -m deepseek-v4-pro

# the thinking model — reasoning streams live as it works
agentty --provider deepseek -m deepseek-reasoner

agentty checks DEEPSEEK_API_KEY first (falling back to OPENAI_API_KEY, or an explicit -k for the session), and knows DeepSeek's endpoint natively — you don't set a URL. --provider and -m persist, so subsequent runs are just agentty. Switch provider live with ^P and model with ^/; both overlays have a search filter, so just start typing deepseek or kimi to jump to it.

Use deepseek-v4-pro for fast day-to-day work and deepseek-reasoner when you want the model to think longer — and with agentty, its chain-of-thought streams live in the same reasoning view as Claude's thinking, Grok's, and Gemini's.

What you get with either

Once you're on Kimi or DeepSeek, this is a real agent, not a chat box:

• Native tool-calling — agentty's full tool suite (read / edit / write files, ripgrep-backed search, structural search, shell, apply-patch, subagents) runs through standard OpenAI-style function calls. Both Kimi K2 and DeepSeek support them, so the entire toolchain works.
• Streaming token output, and live-streamed reasoning for the thinking models — DeepSeek's reasoning_content surfaces through the same event as every other provider's, so it renders identically.
• In-app model picker (^/) populated from the provider's live catalog.
• Cost + token accounting per turn — and both Kimi K2 and DeepSeek are among the cheapest capable models you can drive.
• Everything else agentty does — Smart Mode routing, retrieval, sandboxed shell, one-command SSH air-gap, Zed over ACP — works the same regardless of which model is behind it.

Why agentty for Kimi / DeepSeek users

agentty ships its own C++ OpenAI-compatible transport — SSE streaming, streamed tool_calls reassembled natively, provider differences collapsed to configuration. So a hosted model like Kimi K2 or DeepSeek V4 is a first-class citizen, not a compatibility shim:

• One static binary. No Node, no Python, no Electron. curl … | sh and you have an agent.
• Millisecond cold start. ~18 MB, starts instantly, so the loop that matters — type → model → tool → result — is bound by the model's latency, not the client's.
• Sandboxed by default, one-command SSH air-gap, runs inside Zed over ACP.
• Your choice of model behind one client — swap Kimi ⇄ DeepSeek ⇄ Claude ⇄ GPT ⇄ Gemini ⇄ Grok live with ^P, no restart.

Frequently asked

Does agentty support Kimi K2? Yes — first-class since 0.4.0. agentty login signs you in with your Kimi plan over OAuth (no API key), then agentty --provider kimi runs you on Kimi K2. Streaming and native tool-calling both work.

Does agentty support DeepSeek? Yes. Set DEEPSEEK_API_KEY and run agentty --provider deepseek -m deepseek-v4-pro (or deepseek-reasoner for the thinking model — its reasoning streams live). DeepSeek's endpoint is built in; you don't configure a URL.

Do I need an API key for Kimi? No. Kimi uses OAuth sign-in (agentty login → choose Kimi), the same way ChatGPT and GitHub Copilot sign-in work. DeepSeek uses a DEEPSEEK_API_KEY.

Do file edits and tools work, or just chat? Full tools. agentty sends standard OpenAI-style tools and reassembles streamed tool_calls, so reading/editing files, shell, grep/ripgrep search, apply-patch, and subagents all work on both Kimi K2 and DeepSeek.

Can I switch between Kimi, DeepSeek, and Claude without restarting? Yes. ^P switches provider and ^/ switches model live mid-thread; both have a search filter — type kimi or deepseek to jump straight to it. Your selection persists between runs.

Does the reasoning show for deepseek-reasoner? Yes — DeepSeek's streamed chain-of-thought surfaces in agentty's reasoning view, the same place Claude/Grok/Gemini thinking appears.

How do I install agentty? curl -fsSL https://agentty.org/install.sh | sh on macOS or Linux (Intel and Apple Silicon). Single static binary — no runtime to install.

---

Kimi K2 and DeepSeek give you frontier-class coding at a fraction of the cost. agentty gives you a native, instant, tool-calling agent to drive them from your terminal — sign in with Kimi, key in DeepSeek, and go: curl -fsSL https://agentty.org/install.sh | sh.