A lightweight, local-first AI agent runtime. Your device. Your rules.
Phase: Planning / Pre-development
Date: 2026-02-03
What: Rust-based AI agent runtime that's lightweight, private, and cost-aware.
Why: Current agent frameworks are heavy, expensive, and developer-only. We want something that runs on a Pi, respects your wallet, and leverages existing tools.
How:
- Small Rust core with capability-based security
- MCP-first tool integration (borrow from community)
- SQLite storage with event logging
- Token budgets and cost tracking built-in
All side effects require an explicit capability.
| Document | Purpose |
|---|---|
| docs/VISION.md | Why we're building this |
| docs/SPEC.md | What to build (M0/M1 scope) |
| docs/STYLE.md | How we write code |
| docs/COMMUNITY.md | MCP-first tool strategy |
| docs/ARCHITECTURE.md | Future architecture reference |
- Humans: See CONTRIBUTING.md
- AI Agents: See AGENTS.md
- CLI chat
- SQLite event log
- Capability gate + policy
- LLM streaming (Anthropic)
- Safe defaults
- MCP client integration
- Tool listing + invocation
- Capability enforcement at tool boundary
- Cost tracking
- Memory system (pinned facts + search)
- HTTP API
- Context budgets + compression
- Portability (WASM, mobile)
Not yet implemented. This is planning documentation.
# Future usage
agent init
agent chatApache 2.0 — see LICENSE