In development · no release yet GPLv3, royalty-free or commercial — your choice

runa.

Local-first AI CLI — fit-check a GGUF before you download it, run it locally via ggml / llama.cpp, or hit OpenAI and Anthropic.

$ curl --proto '=https' --tlsv1.2 -LsSf https://github.com/pyrlyn/runa/releases/latest/download/runa-installer.sh | sh

View on GitHub Read the docs

runa has no published release yet, so the installer above works from the first GitHub release. Until then, build from source (see Install below).

The runa mark: a geometric rune stave with an ember glowing at its base, on warm charcoal glass.

See it run

runa, in the terminal.

runa — zsh

$ runa fit hf:unsloth/Qwen3-30B-A3B-GGUF:Q4_K_M --ctx 16384 --kv q8_0
$ runa fit --recommend --use code
$ runa fit --recommend --offline --top 5
$ runa run qwen "explain KV-cache quantization in one paragraph"
$ runa chat qwen --tui
$ runa serve --port 8080
$

$ runa fit hf:unsloth/Qwen3-30B-A3B-GGUF:Q4_K_M --ctx 16384 --kv q8_0
$ runa fit --recommend --use code
$ runa fit --recommend --offline --top 5
$ runa run qwen "explain KV-cache quantization in one paragraph"
$ runa chat qwen --tui
$ runa serve --port 8080

Commands from the runa usage examples.

Overview

What runa does.

runa is one CLI binary that runs AI models locally (GGUF via ggml / llama.cpp) or through the OpenAI and Anthropic APIs. Thinking controls, structured output, tools and runa serve work the same way on both sides.

Before you download a model, runa fit tells you whether it runs on this machine and how fast, and the forecast calibrates to your device as you record runa bench runs. Release binaries are portable CPU builds for macOS arm64, Linux x86_64 and Windows x86_64; Metal, Vulkan and CUDA builds are extra artifacts. No telemetry.

runa is in development: the v0.1.0 tag has no GitHub release yet, so the installer and the Homebrew formula below work from the first release. Until then, build from source.

Features

Built for real workflows.

Fit before fetch

runa fit gives a memory and speed forecast before any download, calibrated by real runa bench runs. runa fit --recommend ranks a curated catalog for your hardware, offline too.
Local and cloud, same controls

Deep thinking (off / on / budget / effort), structured output and tools work the same way for GGUF models and for the OpenAI and Anthropic APIs.
Structured output

Force an answer into a JSON Schema or a GBNF grammar; local models use llama.cpp grammar sampling, so the output always matches.
MCP tools

--mcp starts stdio MCP servers and the model calls them in a loop in run and chat; runa serve supports tool calling.
Media in the loop

Audio and video through native mmproj, ASR (whisper.cpp) or the cloud; runa media probe|video|transcribe.
Serve and stay warm

An OpenAI-compatible HTTP server (runa serve) and a local daemon socket for snappy run / chat; compute modes cpu / gpu / hybrid plus auto placement.
Adaptive memory

Shrinks toward a floor when idle, with a bounded pre-grow when a task is heavy.

Install

Up and running in a minute.

Install

From a GitHub Release (portable CPU archives; Metal / Vulkan / CUDA are extra artifacts):

$ curl --proto '=https' --tlsv1.2 -LsSf https://github.com/pyrlyn/runa/releases/latest/download/runa-installer.sh | sh

Or Homebrew from the formula on that Release:

$ brew install ./runa.rb

Host-tuned local build:

$ RUSTFLAGS='-C target-cpu=native' cargo build --release --features native

Usage

Fit-check a model before downloading it, and find the best catalog models for this machine:

$ runa fit hf:unsloth/Qwen3-30B-A3B-GGUF:Q4_K_M --ctx 16384 --kv q8_0
$ runa fit --recommend --use code
$ runa fit --recommend --offline --top 5

More examples below.

Usage examples

Real commands, straight from the docs.

Run, chat, or serve an OpenAI-compatible HTTP API:

$ runa run qwen "explain KV-cache quantization in one paragraph"
$ runa chat qwen --tui
$ runa serve --port 8080

Calibrate speed predictions with a measured run:

$ runa bench qwen2.gguf --mode cpu --pp 512 --tg 128
$ runa fit qwen2.gguf --json

Cap reasoning tokens, or let the model call MCP tools:

$ runa run qwen3-8b.gguf "What is 2+2?" --think-budget 32 --max-tokens 48 --json
$ runa run qwen3 "What is the weather in Paris?" --mcp 'python3 weather_server.py'

Transcribe audio with whisper.cpp:

$ runa media transcribe --model base --lang auto clip.wav

Links

Go deeper.

License

You can use runa under any of the following licenses, at your choice:

GNU GPLv3 — free for open-source applications on any platform, including embedded systems.
Royalty-free license — free for proprietary desktop, mobile and web applications, as long as you disclose that your application uses the project. Embedded systems are not covered.
Commercial license — for proprietary applications, including embedded systems, without the attribution requirement.

Pricing

Pro is a placeholder

Free forever. Pro when you need it.

Free · open source

GPLv3, royalty-free or commercial — your choice

$0forever

Fit check before any download
Local GGUF and cloud APIs, same controls
OpenAI-compatible runa serve

Install free

runa Pro

Placeholder

$9 / moper seat

Team features
Hosted dashboard
Priority support

Placeholder — the Pro waitlist is not open yet.

Try runa today.

Free and open source. Star it, file an issue, or send a pull request.

Read the docs github.com/pyrlyn/runa

runa.

Fit before fetch

Local and cloud, same controls

Structured output

MCP tools

Media in the loop

Serve and stay warm

Adaptive memory

Install

Usage

License

Free · open source

runa Pro

Try runa today.