fx is a coding agent CLI running as a Zig-built native binary, released in August 2026 by Vercel Labs, Vercel’s experimental projects division. It commits to the lightness advertised on its official site ─ roughly 10 microseconds of cold start and a binary in the 6MB range ─ and is designed primarily for embedding into agent sandboxes and for research-oriented harness experimentation. It is open source (Apache-2.0) and supports Mac and Linux.
Key Features
- Lightweight and fast to start: Cold start for the native binary is about 10 microseconds according to the official site. Binary size is stated as 6.19MiB on the official site, while the GitHub README lists 7.8MiB ─ the figure differs by source. Memory usage for a single process also stays in the single-digit MB range, a design aimed at embedding agents into sandboxes in large numbers.
- Multi-provider support: The default path is usage via the Vercel AI Gateway, but you can also connect a ChatGPT Plus/Pro/Business/Enterprise/Edu subscription with
fx login codex, or a SuperGrok/X Premium subscription withfx login grok. AI Gateway API keys and local inference endpoints work as well. - MCP, skills, and subagents: Beyond connecting to external tool servers as an MCP client, fx can load reusable instructions (skills) and offers a subagent mechanism for delegating and running independent tasks in parallel.
- Embedding and ACP support: In addition to native binaries, fx ships WebAssembly builds (fx-core.wasm / fx-term.wasm), so you can embed it in a Node SDK or in the browser. It can also run as an Agent Client Protocol (ACP) server, callable from external clients such as editors.
- Privacy-first design: fx sends no product telemetry and never auto-uploads your workspace, Git history, or session files. It also supports fully hermetic operation with outbound traffic blocked (local inference only).
- Graduated permission modes: In the default auto mode, fx asks for approval only on risky operations such as file writes and command execution. You can adjust modes and individual rules with
/permissions, and a yolo mode that disables all checks is available as well (use at your own risk).
Pricing
| Plan | Monthly price | Key details |
|---|---|---|
| CLI itself | Free (OSS) | Apache-2.0 license. fx itself has no usage caps or paid tiers |
Pricing reflects information as of September 2026. fx is a free open-source CLI; costs arise on the side of the inference provider you connect to (metered billing on the Vercel AI Gateway, or your existing ChatGPT/Grok subscription). The official site has no page equivalent to /pricing ─ see Usage and costs for details.
Pros and Cons
✅ Pros
- Runtime cost per instance is small, so resource footprint stays manageable even when you spin up an agent per CI job or per task
- You can use your existing ChatGPT or Grok subscription as the inference source without extra charges, so you can try it without signing up for a new API contract
- Since it sends no telemetry and runs with outbound traffic blocked, it is easier to make the case for adoption in environments where code cannot leave the organization
- With WebAssembly builds and ACP server support, you can embed it as the internal engine of your own tools or editors, not just use it as a CLI
⚠️ Cons
- Status is experimental, and the project explicitly states that frequent breaking changes may occur
- Supported operating systems are Mac and Linux only; Windows/WSL is not supported
- As a new project released in August 2026, its production track record and third-party ecosystem are thin compared to established tools like Claude Code
- Because the design deliberately avoids a heavyweight TUI, users expecting a rich IDE-like interface may find the feature set lacking
Comparison with Similar Services
| Service | How it differs from fx |
|---|---|
| Claude Code | Anthropic’s official terminal-based coding agent. Richer in features and ecosystem, but the binary is far larger than fx’s |
| OpenAI Codex CLI | OpenAI’s official terminal-based coding agent. fx can use Codex’s ChatGPT subscription as one of its connectable providers |
| Gemini CLI | Google’s open-source coding agent. More capable and feature-rich than fx, but heavier in dependencies and binary size |
| opencode | An open-source coding agent CLI written in TypeScript. fx is a Zig-built native binary that prioritizes startup speed, low memory use, and embeddability |
Who It’s For
- People who want to embed a low-memory, low-latency coding agent into agent sandboxes or massively parallel execution environments
- People who want to inspect and modify the internal design of an agent harness (system prompts, tool configuration, and so on) for research purposes
- People who need hermetic code execution with the network cut off, or offline use on highly confidential repositories
- People who want to connect fx to external clients such as editors via an ACP server and fold its agent into an existing workflow
Summary
The deciding factor for adoption comes down to “how many agents will you run?” If you use one carefully, established tools like Claude Code are enough; if you run many in parallel across CI or sandboxes, the difference in startup cost starts to matter. That said, the status is experimental, and between the lack of Windows support and the thin production track record, checking the latest documentation and changelog before any production adoption is essential.