recursive-oss/rec

rec is our custom inference engine for running Recursive's language models locally on your machine. It provides lightning-fast inference for single-session usage. Compared to llama.cpp, rec delivers 1.4–2.6× the token throughput across the context lengths shown below on M4 Pro. Use it with pi, OpenCode, or connect your own client through the local API.

_{Flint on Apple M4 Pro (48 GB), 4-bit weights, 1K–252K-token prompts and 1,024 output tokens. Five runs per context; prefill excluded.}

• Apple Silicon Mac with an M4 Pro, M5 Pro, or M5 Max
• At least 48 GB unified memory
• About 30 GB free disk space
• Node.js 22 and npm for pi

Install using:

curl -fsSL https://raw.githubusercontent.com/recursive-oss/rec/main/install.sh | bash

This installs rec in ~/.local/bin and adds it to your PATH for new terminals. All other data lives under ~/.local/share/rec.

We find the best performance from our models when paired with the pi harness. For convenience, rec provides pre-configured recipes for running pi and OpenCode harnesses. Run them using:

rec launch pi  # preferred
rec launch opencode

⚠️ rec is optimized for single-session usage. To run multiple agents at a time, see our documentation for other ways to run our models.

This starts our inference server and connects to it automatically. The first run downloads the model and harness. The model server continues to run after you close the terminal. Run rec stop to stop it.

These harnesses are kept separate from anything on your system. Use rec exec pi|opencode ARGS to run their own commands, for example:

rec exec pi install npm:PACKAGE

rec exposes OpenAI and Anthropic-compatible chat-completion APIs. Start the server:

rec serve

Keep this terminal open. When flint is ready. appears, connect a client from another terminal.

To use your own install of pi or OpenCode, run:

rec launch --args pi
rec launch --args opencode

Each command prints the environment and command line that rec launch uses, instead of running it. Replace the harness path with the path to your own harness. The command also points the harness at rec's own settings and sessions; remove those variables to keep yours.

To send a request directly to the OpenAI-compatible API:

curl -fsS http://127.0.0.1:11500/v1/chat/completions \
  -H "Authorization: Bearer $(rec auth token)" \
  -H 'Content-Type: application/json' \
  -d '{"model":"flint","messages":[{"role":"user","content":"Say hello."}],"max_tokens":128}'

The response contains generated text in choices. For other clients, use base URL http://127.0.0.1:11500/v1, model flint, and the API key from rec auth token.

The Anthropic-compatible API is available at http://127.0.0.1:11500/v1/messages. Point ANTHROPIC_BASE_URL at http://127.0.0.1:11500 and send the token from rec auth token as a bearer token or as x-api-key. The model is flint.

Run rec COMMAND --help for a command's options.

• Questions, bugs, and feature requests: GitHub Issues.
• Contribution guide and local checks: CONTRIBUTING.md.
• Community guidelines: Code of conduct.
• Security vulnerabilities: Private reporting instructions.
• License: Apache 2.0.