API reference

Signatures and docstrings for the tulip package, rendered from tulip-agents 2.16.0, the SDK version this site is built against. Each page covers one area of the framework. For what a piece is for and how it fits with the rest, start with the guide linked beside it, then come back here for the exact signature.

The everyday names import from the package root. Everything else imports from the module its page names:

from tulip import Agent, AgentConfig, StateGraph, tool
from tulip.core.termination import MaxIterations, ToolCalled

Agents and runs

Reference	What it documents	Guide
Agent	Agent, AgentConfig, AgentResult and AgentState: configure an agent, run it, and read back a typed result	Agent, The agent loop
Termination	Stop conditions such as MaxIterations, TimeLimit and ToolCalled, composed into one rule and set on AgentConfig.termination	Termination
Hooks	HookProvider and HookRegistry, which dispatches providers in priority order; the hook event types; and the built-in logging, telemetry, retry, guardrail and steering hooks	Hooks
Events	The typed, frozen TulipEvent models that async for event in agent.run(...) yields	Events
Streaming	Handlers that send the event stream to a terminal, a Server-Sent Events response or a buffer, and structured-output streaming	Streaming, Graph streaming
Core primitives	The shared types underneath: settings, protocols, messages, the error hierarchy, and the Command and interrupt helpers for graph control flow	State, Errors

Tools and capabilities

Reference	What it documents	Guide
Tools	The @tool decorator, which builds a tool's schema from the function's type hints and docstring; ToolContext, the registry and the executors	Tools
Capability providers	Web search, web fetch, image generation and speech providers. Set one on AgentConfig and the agent registers a matching tool	Multi-modal providers, Tools
Integrations	MCP in both directions: serve a Tulip agent to any MCP client, or turn MCP tools into Tulip tools. Also use_aws, which refuses any AWS operation whose name does not start with a read verb such as Describe, List or Get	MCP
Skills	Skill and SkillsPlugin: SKILL.md instruction bundles the agent loads on demand, attached with AgentConfig.skills	Skills
Playbooks	Declared step sequences with expected tools. Set AgentConfig.playbook and a PlaybookEnforcerHook checks each tool call against the current step	Playbooks

Models and reasoning

Reference	What it documents	Guide
Models	get_model and the provider registry, where the part of a model string before the colon picks the provider; the ModelProtocol contract; and the native provider classes	Model providers
Reasoning	Reflector, which evaluates the agent's progress after each iteration; GroundingEvaluator, which checks claims against the evidence the tools gathered; CausalChain; and GSAR's thresholds and results	Reasoning, GSAR

Memory and retrieval

Reference	What it documents	Guide
Memory	Conversation managers that keep, window or summarize the history within a run; the cross-thread BaseStore; and the long-term memory manager, which decides what to keep from a finished session	Conversation management, Long-term memory
Checkpointers	BaseCheckpointer and its storage backends. Prior turns are reloaded only when the agent has a checkpointer and you pass a thread_id	Checkpointers
RAG	RAGRetriever over pluggable embedders and vector stores, rerankers, multimodal processing, and helpers that expose retrieval as a tool	RAG

Multi-agent

Reference	What it documents	Guide
Multi-agent	Pipelines (SequentialPipeline, ParallelPipeline, LoopAgent), an orchestrator with specialists, swarms, handoffs, and the StateGraph workflow	Multi-agent workflows
DeepAgent	create_deepagent, a research-shaped Agent with optional filesystem, todo and subagent tools; and create_research_workflow, a StateGraph that scores its summary's grounding and rewrites or replans, up to set limits, when the score is below threshold	DeepAgent
A2A	A2AServer and A2AClient: one agent calls another over HTTP with the Agent-to-Agent protocol. The server accepts both the v1 wire format and the earlier one	A2A — Agent-to-Agent

Serving, observing and testing

Reference	What it documents	Guide
Agent server	AgentServer, which wraps an agent as a FastAPI app with /invoke, /stream, /threads/{thread_id} and /health routes; and GraphRunnable, for serving a compiled StateGraph	Agent Server
Observability	The in-process EventBus, which fans events out to subscribers by run id; the emit() helpers; run context; and EventBusHook, which republishes an agent's hook events onto the bus	Observability, SSE event catalogue
Evaluation	EvalCase, EvalRunner and EvalReport for test suites; LLMJudge for answers with no single right string; check_trajectory for the order tools were called in; and graph evaluation targets	Evaluation, Testing agents

Control

Reference	What it documents	Guide
Control	admit(), which checks an action against a ControlPolicy you write, before the action runs: allow, hold for a person, or deny (API values allow, require_human, deny). gate_tool puts the same check in front of a tool, approval stores keep held calls until someone decides, and AuditTrail records each decision when you pass one. Only the calls you route through the gate are checked	Control-layer architecture, Writing a policy that holds

Security domain

Reference	What it documents	Guide
Security	One domain built on the framework: red-team probes and jobs, grounded findings that come back as evidence or an Abstention, verification skeptics, and MITRE ATLAS and OWASP taxonomy tags. Import the admission gate from tulip.control, not from here	Agentic AI-security, Security layer

Deprecated

ReAct loop documents tulip.loop, a second ReAct implementation that Agent has never used. It is scheduled for removal in 3.0.0, and each name you import from it emits TulipDeprecationWarning. Use Agent instead.