Syscity is an agent system — a runtime that lets AI agents act on your computer. Unlike chatbots that only read and write text, Syscity agents can control your desktop, execute code, operate your browser, and manage your files.
Traditional AI lives inside a browser tab. Syscity lives inside your machine.
One agent runtime, every device. Run the same local agent on macOS, Windows, Linux, iOS, and Android — your data, memory, and tools stay on your machine. What leaves it goes where you point it: your model provider, and any chat channel or MCP server you connect.
For developers who want to build LLM-powered automation. For power users who want AI to control their desktop, not just chat.
Install
# macOS / Linux
curl -sSL https://syscity.net/install.sh | bash
syscity setup picks a model provider; syscity start serves the Web UI at
http://127.0.0.1:18080. To build from source, see docs/build.md.
Benchmarks
Scores come from each benchmark's official evaluator, never a homegrown
scorer. Model: deepseek-v4-flash throughout.
| Benchmark | What it measures | Score |
|---|
BFCL bfcl-eval 2026.3.23 | the model plus the OpenAI-compatible wire (raw mode) | 80.12% (673/840) over two runs (78.81% then 80.12%, a difference inside the noise floor) |
| GAIA validation Level 1 | syscity's own agent loop, end to end | 90.6% (144/159 trials); 70.4% → 78.0% → 90.6% over three runs (the last bundles two changes — see the notes) |
Finer-grained tables, the method behind each number, the measured noise floor,
reproduction steps and the failure breakdown are in
docs/benchmark.md. τ²-bench is next.
Overview
Platforms and mobile apps, why Syscity exists, the architecture diagram, the
action/cognition tool lists, the full quick start, macOS desktop control and the
configuration reference: docs/overview.md.
Documentation
License
Apache-2.0
Contributing
PRs are welcome! Check out the issues or join the discussion on Discord.
