SVTPrivateAI
The model proposes. The runtime decides.
A local-first Python runtime that runs AI agents under scoped capabilities, a deterministic policy engine, human approvals and a hash-chained audit log.
Agents are declared in YAML with scope-aware capabilities — read wide, write only where needed — and every tool call goes through one controlled path: capability check, policy evaluation, an optional human approval bound to that exact call, then a tamper-evident audit record that names the rule that acted.
Providers include Anthropic, OpenAI-compatible endpoints and a deterministic mock used by default, so it runs without an API key. MCP servers are treated as untrusted extensions, and secrets are resolved inside tools, never placed in model context. The `svt` CLI checks, explains and audits every decision.
- Type: AI agent security runtime (CLI)
- Status: In development — Version 0.1 (alpha): the foundation is complete and tested; not yet hardened for multi-tenant or internet-facing use.
- Code: simonvildonelohim/svtprivateai (private repository)
- Year: 2026
Built
- Scoped capabilities that can only narrow
- Deterministic, deny-by-default policy engine
- Human approvals bound to one exact call
- Hash-chained, tamper-evident audit log
- Secret broker: secrets never reach the model
- Path jail, SSRF checks, no shell
- Anthropic, OpenAI-compatible and mock providers
- `svt` CLI: doctor, policy check, approvals, audit
Technology
Python, Pydantic, Typer, Rich, httpx, SQLite, OpenTelemetry, MCP, pytest, mypy
