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. 
## Requirements
- 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
Install using:
 [code] 
This installs `rec` in `~/.local/bin` and adds it to your PATH for new
terminals. All other data lives under `~/.local/share/rec`.
## Usage
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:
 [code] 
⚠️ 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:
 [code] 
### Use your own harness
`rec` exposes OpenAI and Anthropic-compatible chat-completion APIs.
Start the server:
 [code] 
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:
 [code] 
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:
 [code] 
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`.
## Commands
| Command | What it does |
|---|---|
| `rec launch pi\|opencode` | Start the server if needed and open a harness |
| `rec exec pi\|opencode ARGS` | Run a harness command, such as installing a plugin |
| `rec serve` | Start the model server in this terminal |
| `rec status` | Show whether the server is running |
| `rec stop` | Stop the server |
| `rec auth token` | Print the local API key |
| `rec pull` | Download the model without starting the server |
| `r