01Lightweight
Small footprint. Fast start.
One native Go binary, no Node.js or Python runtime. A ~12 MB download and ~2.3k tokens of harness context leave more resources for your task.
Open-source terminal agent
Minimal overhead.
Minimal overhead.
Maximum agent.
One native binary. Your choice of model. Shape agent behavior and coordinate work across agents, with control at every step.
Start with San
brew install genai-io/san/san
macOS, Linux & Windows · Apache 2.0
Releases and downloads
A quick introduction
Meet San in 70 seconds.
From choosing a model to working across agents. The original comic tour, at your pace. Model names and limits are illustrative; performance figures are version-specific.
Why San
Less framework. More room to work.
Lightweight by design, adaptable to your workflow. The essentials for capable agents, without prescribing how you use them.
02Your harness
Your model. Your way of working.
A model-agnostic runtime with configurable system prompts, personas, skills, and tools. Shape agent behavior without adopting someone else's workflow.
03Collaboration
One task. More than one agent.
Delegate to subagents, exchange messages across sessions, and assemble dynamic workflows. Split the work without losing the thread.
/group
04Control & inspection
Decide what runs. See what happened.
Configure execution permissions and inspect model context and tool calls. Set the agent's scope, then review the run rather than guessing.
Shift+Tabsan inspector
Benchmark
A smaller runtime, measured.
Against Claude Code on Apple Silicon, with the same model (claude-sonnet-4-6) and tasks. Results are specific to the versions below.
| Metric | San | Claude Code | Comparison |
|---|---|---|---|
| Download size | 12 MB | 63 MB + Node 112 MB | 5× smaller¹ |
| Disk footprint | 38 MB | 175 MB | 4.6× smaller |
Startup time (version) | ~0.01s | ~0.20s | 20× faster |
| Startup memory | ~32 MB | ~189 MB | 5.8× less |
| Simple task | ~2.4s / 39 MB | ~10.4s / 286 MB | 4.3× faster |
| Tool-use task | ~3.3s / 39 MB | ~26.0s / 285 MB | 7.9× faster |
| Harness context² | ~2.3k tokens | ~20.9k tokens | ~9× leaner |
Task figures are wall time / peak RSS in print mode; network and API latency vary. Startup measures the version command, not time to an interactive prompt. ¹ Download comparison excludes the additional Node runtime. ² Harness context is system prompt + tool schemas on an empty first turn, measured separately on San v1.22.0 vs Claude Code v2.1.220. All other rows are from the San v1.13.2 / Claude Code v2.1.112 run. Method and reproduction →
Community
Build with us.
Share workflows, ask questions, and help shape San.
关注公众号「极客外传」· 回复 san 或 三 入群