Model-agnostic
One Provider interface over Anthropic, OpenAI, Google, OpenRouter, Ollama and anything
OpenAI-compatible. Swap models with one line.
intellibrix
Compose agents from small, typed, testable bricks. Any model, any provider, human approvals built in.
Typed end to end
Tools and Programs use Zod 4 schemas (via Standard Schema). Inputs are validated before they run, and structured output is validated before you see it.
Safe by default
Writes, shell and code execution need approval. Outside content is tagged untrusted, and file tools are jailed to a workspace.
Testable offline
MockProvider, record/replay and conformance suites make agents deterministic to test, with no
network or API keys.
MCP both ways
Mount remote MCP tools into a brick, or serve your bricks to Claude Desktop, Cursor and others.
Multi-agent
Supervisors, routers, pipelines and handoffs, all built from the same bricks.