Claude Code and Cursor are full-featured, but sometimes you don't want a chat window — you want a binary that starts before you finish pressing Enter and disappears into a script, a CI job, or a sandbox. fx (from Vercel Labs) and agentty are both built for exactly that: no chat UI, no heavyweight runtime, just a coding agent that starts in milliseconds and talks to whichever AI provider you point it at. They solve the same problem from different angles — one is built to embed inside other tools and pipelines, the other is built to run fully offline on your own machine.
The short version
| fx | agentty | |
|---|---|---|
| Binary size | Under 8 MB | ~16.7 MB |
| Cold start | Microseconds | ~3 milliseconds |
| License | Apache-2.0, backed by Vercel Labs | MIT, independent project |
| Model support | Multiple providers, cloud or local | Claude, GPT, Groq, DeepSeek, Grok, Gemini, Mistral, or fully local Ollama |
| Standout feature | Embeds inside other tools via WebAssembly, skills, plugins and MCP | Built-in local RAG (BM25 + embeddings + GraphRAG) for navigating a codebase's docs |
| Built for | CI pipelines, scripts, sandboxes — spawned by machines | A developer's own terminal, including fully offline use |
| Maturity | Explicitly experimental, just open-sourced | Young (under 600 GitHub stars), MIT |
fx — built to disappear into your infrastructure, not to chat
fx is Vercel Labs' tiny, open-source coding agent: an 8 MB native binary written in Zig that starts instantly and is meant to be embedded inside other tools rather than run as a standalone chat app. It's model-agnostic, works with several AI providers whether local or cloud, and supports WebAssembly so it can be dropped into other products, plus extensibility through skills, plugins and MCP.
Who it's for: DevOps teams and infra builders who want to spawn an AI coding agent programmatically inside a CI pipeline, a script, or a sandbox — or embed one directly inside another product.
Pricing: Free and open source (Apache-2.0). No hosted cost; you bring your own model provider.
Honest limits: its own creators call it explicitly experimental, permission handling for autonomous execution needs close attention, provider and authentication integrations are more limited than mature agents, and there's very little production track record since it was just open-sourced.
agentty — a single binary that never has to touch the network
agentty is a single-binary terminal coding agent with a roughly 3-millisecond cold start and no runtime dependencies — no Node, no Python. It works with Claude, GPT, Groq, DeepSeek, Grok, Gemini and Mistral, or runs completely locally through Ollama with no API key at all, and it ships with a built-in local retrieval system (BM25, embeddings and GraphRAG) so it can navigate a project's documentation on its own.
Who it's for: developers who want an instant-start terminal coding agent that isn't locked to one model provider, including a fully local, API-key-free setup, plus built-in retrieval over a codebase's docs.
Pricing: Free and open source (MIT). Provider API keys are billed separately by that provider, or $0 if you run it fully local via Ollama.
Honest limits: a small ecosystem so far (under 600 GitHub stars, versus tens of thousands for Claude Code or Aider), building from source requires a C++26 toolchain (GCC 14+, Clang or MSVC), sandboxed shell execution isn't available on non-rooted Android, and it has fewer third-party integrations than established agents.
Pick fx if…
…you're building infrastructure — a CI pipeline, an internal tool, a fleet of sandboxed agents — and need something a machine can spawn on demand and embed via WebAssembly or MCP, rather than a human typing prompts into it directly.
Pick agentty if…
…you want one fast binary for your own terminal, the freedom to switch between Claude, GPT, Groq, DeepSeek or a fully offline Ollama model without changing tools, and built-in local retrieval over your project's docs.
Both are genuinely new — weeks old, open source, and free — so neither has the track record of Claude Code or Aider yet. If you just want the fastest way to try one today, agentty's Ollama-only mode costs nothing and never leaves your machine; fx's edge shows up once you want that agent living inside something else you're building, not typed into directly.