Guides
Use these guides when you have a specific runtime task to complete. They assume you already have a RuntimeSpec or have worked through the Quickstart.
Use these guides when you have a specific runtime task to complete. They assume
you already have a RuntimeSpec or have worked through the Quickstart.
Require approvals before execution
Use Approvals when plans or tools should pause for a host application decision before the runtime continues.
Configure a data environment
Use Data Environment to connect Rust-owned storage, catalog, target database, and knowledge dependencies through runtime config.
Profile data and export a catalog
Use Profiling and Catalog Export to profile a read-only target database with the CLI, persist the resulting catalog, and export a portable artifact your runtime consumes through the data environment.
Attach knowledge and documents
Use Knowledge and Documents to index source material and expose retrieval-backed context to agents through the built-in knowledge tools.
Expose tools over MCP
Use Expose Tools Over MCP to serve runtime tools to MCP-aware clients over stdio or Streamable HTTP.
Run the local Studio UI
Use Studio to chat with your agents in a local browser interface, browse attached data sources, create tests, run evals, and inspect runs and traces.
Debug system prompts
Use Debugging System Prompts to find the stage where an
agent's system prompt went missing or stale, from layered_prompt(...) down
to the provider call. For prompt authoring, see
Prompts.
Test an agent without provider credentials
Use Testing to run deterministic and scripted interpreters, exercise approval flows, and assert emitted events without calling a live model provider.
Write evals
Use Evals to score recorded samples, run runtime-backed evals, and replay scripted coordinator flows through real tools and plan execution.
Judge final responses with a model
Use Judge Evals to score final response quality with a rubric-driven LLM judge, weight judges against deterministic scorers, and read verdicts and traces from eval artifacts.
Stream runtime events
Use Streaming Events to adapt runtime event dictionaries into an HTTP or application-specific event stream.
Runtime
define_runtime(...) produces a validated RuntimeSpec; create_runtime(...) hands the spec to the Rust runtime (flowai-runtime) and returns a native Runtime handle.
Approvals
The Flow AI runtime can pause mid-stream to ask the host application to approve a proposed plan or a pending tool call. Python supplies the policy and the response; the actual...