Skip to main content
rec is the command-line tool for Recursive’s local inference engine, optimized to run Flint fast on your Mac. It downloads the model, starts the local server, and connects it to your coding harness. Launch Pi or OpenCode in one command, or connect your own client to the local API. You can also run Flint locally through Ollama, Unsloth, LM Studio, or llama.cpp.
NVIDIA RTX Spark support for rec is coming soon. The setup instructions below apply to supported Macs.

Mac requirements

  • Apple Silicon Mac with an M4 Pro, M5 Pro, or M5 Max and Metal support.
  • At least 48 GB unified memory.
  • About 30 GB free disk space for the model and runtime.
  • Node.js 22 and npm to launch Pi or OpenCode.
rec checks your chip and memory before downloading Flint. It will not download the model on a Mac with an unsupported chip or less than 48 GB of unified memory.

Install rec

Run the installer from the rec repository in Bash or Zsh:
The installer downloads the latest rec release, verifies its checksum and size, checks your Mac, and installs rec in ~/.local/bin. It adds that directory to your shell’s PATH for new terminals. The model downloads on first use. Open a new terminal, or update the current terminal’s PATH and check the installation:
To update rec, run the installer again. It stops a running model server before updating. We recommend Pi for local Flint coding. Its lighter harness generally adds less overhead. Informal usage reports suggest roughly 1.5× faster workflows than OpenCode, though results vary by task, hardware, and configuration. The estimate describes the overall coding workflow.

Launch Pi

After installation, run this from your project directory:
The launcher starts or reuses the local server and supplies authentication, model configuration, and runtime settings. The first launch downloads the model and the matching Pi version. No separate Pi installation, manual provider configuration, or server startup is needed.

Launch OpenCode

If you prefer OpenCode, run this from your project directory:
rec provides the same automatic server and connection setup with its matching OpenCode version.
Keep the launcher’s bundled sampling and output settings. The model server continues running after you close either harness or its terminal. Run rec stop when you are finished.

Connect another client

rec serve exposes an OpenAI-compatible Chat Completions API for existing clients. The Pi and OpenCode launchers are optional: they automate server startup and client configuration. You can connect an editor extension, native VS Code Chat, or your own application directly. Start the server in any terminal, including VS Code’s integrated terminal:
On first use, rec downloads the matching model and runtime. Later runs reuse them. Wait for flint is ready and keep this terminal open while using your client. Treat the local token as a credential. Keep it in your client’s secret field and out of shared configuration files.

Cline in VS Code or Desktop

Open Cline Settings → API Configuration and choose OpenAI Compatible. Set the Base URL to http://127.0.0.1:11500/v1, Model ID to flint, and paste the output of rec auth token into the API-key field. Start a new task with Flint. See the Cline guide.

Native VS Code Chat

Run Chat: Manage Language Models, choose Add Models → Custom Endpoint, and select Chat Completions. Use model flint, the full endpoint http://127.0.0.1:11500/v1/chat/completions, and your rec token at the credential prompt. The VS Code guide includes a complete configuration.

Direct API request

With the server running, send a request from another terminal:
The response uses the Chat Completions format. For coding harnesses, also verify file edits and tool execution. Match client token budgets to the running server. These localhost URLs assume the client runs on the same Mac; a remote extension or container has a different localhost.

Manage rec

Run rec COMMAND --help for a command’s options. The engine handles one generation at a time; overlapping requests can receive HTTP 429.

Storage and uninstall

By default, rec stores its installation, models, settings, and server state under ~/.local/share/rec, with a ~/.local/bin/rec symlink. If you set an absolute XDG_DATA_HOME, rec uses $XDG_DATA_HOME/rec for its data directory. To remove rec:
The command shows what it will remove and asks for confirmation. After confirmation, it stops the server, removes rec’s managed command links, and deletes its data directory, including downloaded models and rec settings. The PATH entry in your shell configuration remains.