Which feature do I need?

Answer the questions top to bottom. Each leaf is a page. Most apps need the first three rows only: a registry, an executor, and the surface that calls it.

flowchart TD
  A{"Do you own the function's code?"} -- yes --> T1["@registry.tool"]
  A -- "no: a library, a class, a service" --> T2["registry.wrap(fn) · import_openapi · import_mcp"]
  T1 --> B{"Who calls the tools?"}
  T2 --> B
  B -- "your code" --> C1["Executor.run / run_call"]
  B -- "a model API" --> C2["renderers + run_tool_loop"]
  B -- "an agent framework" --> C3["integrations.<framework>"]
  B -- "another language or process" --> C4["HTTP server · MCP"]
  C1 --> D{"Can a call do harm?"}
  C2 --> D
  C3 --> D
  C4 --> D
  D -- "yes, decide now" --> D1["confirm= + CallbackApproval"]
  D -- "yes, a person decides later" --> D2["confirm= + DeferredApproval + a store"]
  D -- no --> E{"Does it cost money?"}
  D1 --> E
  D2 --> E
  E -- yes --> E1["Ledger + Prices + BudgetHook"]
  E -- no --> F{"Same call again and again?"}
  E1 --> F
  F -- yes --> F1["cache_ttl + cache_scope · executor.model(cache_ttl=)"]
  F -- no --> G{"Several steps?"}
  F1 --> G
  G -- "independent calls" --> G1["run_many · as_completed"]
  G -- "steps that depend on each other" --> G2["run_plan"]
  G -- "branches, loops, pauses" --> G3["Flow + FlowRunner (or compile to LangGraph / ADK)"]
  G -- "one call at a time" --> H["done"]

Where each leaf is explained

Leaf	Page
@registry.tool	Quickstart, Tools
registry.wrap, importers	Wrap existing code, OpenAPI, MCP
Executor.run, run_call	Executor
renderers, run_tool_loop	OpenAI, Anthropic, Tool loop
a framework	Integrations, support matrix
another language	HTTP server, MCP
a non-Python framework or a hosted agent	Connect anywhere
CallbackApproval, DeferredApproval	Human in the loop, Approval over Slack or email
ledger and budgets	Budgets, Cap spend per user
caching	Caching
batches and plans	Parallel retrieval
flows	First flow, Flows
sync code (Flask, Celery, scripts)	Sync code and backends
a queue, a Lambda or a pipeline	Queues, Serverless, Pipelines
see what ran: flow graph, timeline, approvals (any framework)	See your runs

Always on

Whatever you pick, every call through the executor gets these:

You get	Without asking
argument validation	before the tool runs; errors safe to send back to the model
typed errors with stable kinds	Errors
events for tracing	Observability
a denied call when a confirm tool has no approval port	Security