# Docs

- Guides: Build, run and test agents

  - **Get started**
  - [Introduction](/docs): An agent framework for TypeScript and Python, built on an append-only event log.
  - [Quickstart](/docs/quickstart): Run your first agent with no API key, then switch to a real model.
  - [Installation](/docs/installation): Run threads from source in TypeScript or Python.
  - [How it works](/docs/how-it-works): A short tour of the append-only log behind every thread.
  - [Comparison](/docs/comparison): How threads compares to the OpenAI Agents SDK, Pydantic AI, LangChain Deep Agents, Strands Agents and the Claude Agent SDK.

  - **Agents & tools**
  - [Agents](/docs/agents/agents): Define an agent once, run it as many times as you like, and handle every outcome as a plain value.
  - [Tools](/docs/agents/tools): Give an agent your own functions, with typed input and a clear promise about what happens if a call is interrupted.
  - [Models](/docs/agents/models): Connect Anthropic, OpenAI or any other provider, with retries and fallbacks handled for you.
  - [Structured output](/docs/agents/structured-output): Get a typed, validated value back from a run instead of free text.
  - [Built-in tools](/docs/agents/built-in-tools): Shell, files, web, git, code intelligence and planning tools, switched on by one option each.
  - [Skills](/docs/agents/skills): Give an agent detailed playbooks it loads only when it needs them, so the prompt stays small.
  - [MCP](/docs/agents/mcp): Connect any Model Context Protocol server in one line, with its credentials kept on your side.

  - **Multi-agent**
  - [Subagents](/docs/multi-agent/subagents): Let an agent delegate a task to another agent that runs in its own thread and reports back.
  - [Handoffs](/docs/multi-agent/handoffs): Pass the conversation to a specialist agent that takes it from there.
  - [Teams](/docs/multi-agent/teams): A lead agent and its subagents coordinate through a shared task list and messages.

  - **Sandboxes**
  - [Sandboxes](/docs/sandboxes/overview): Give an agent its own isolated machine to run commands and edit files, with no internet and no credentials by default.
  - [E2B](/docs/sandboxes/e2b): Run an agent's commands and files in an E2B cloud sandbox.
  - [Daytona](/docs/sandboxes/daytona): Run an agent's commands and files in a Daytona sandbox, with snapshots for forking.
  - [Modal](/docs/sandboxes/modal): Run an agent's commands and files in a Modal sandbox (Python).

  - **Channels & host**
  - [Host server](/docs/host/overview): Serve your agents over HTTP, chat channels and schedules from one small server.
  - [Slack](/docs/host/slack): Let people talk to your agent in Slack channels, threads and DMs, and approve its actions with a button.
  - [WhatsApp](/docs/host/whatsapp): Put your agent on a WhatsApp Business number through the WhatsApp Cloud API.
  - [GitHub](/docs/host/github): Let your agent answer issues and pull requests, with approvals as plain comment replies.
  - [Schedules](/docs/host/schedules): Run an agent on a cron schedule, in any time zone, without double runs.
  - [HTTP API](/docs/host/http-api): Start runs, stream their events and control threads from any language over HTTP.

  - **Memory & knowledge**
  - [Memory](/docs/memory/memory): Let an agent remember facts about each user across runs, stored locally or in Supermemory or Zep.
  - [Knowledge base](/docs/memory/knowledge): Let an agent search your documents and cite them, with the exact versions recorded for every run.

  - **Control**
  - [Hooks](/docs/control/hooks): Run your own code at every step of the agent loop to guard, steer or watch it.
  - [Permissions & approvals](/docs/control/permissions): Decide which tool calls run on their own, which wait for a person, and which never run.
  - [Human-in-the-loop](/docs/control/human-in-the-loop): Pause a run for a person's decision, then pick up exactly where it stopped.
  - [Budgets](/docs/control/budgets): Cap tokens, requests, turns, time and money for a thread and everything it starts.

  - **Production**
  - [Durability & crash safety](/docs/production/durability): Pick up after a crash where the agent left off, without silently repeating a charge, an email or any other side effect.
  - [Deploying the host](/docs/production/deploying): Run the host in production with a persistent store, secrets from the environment and safe restarts.
  - [CLI](/docs/production/cli): Serve a host and inspect, move and clean up threads from the command line.

  - **Evals & testing**
  - [Timeline](/docs/evals/timeline): See every step an agent took: what it was asked, what the model saw and said, and every tool call and result.
  - [Fork](/docs/evals/fork): Branch a real thread from a saved point and try a different input, model or prompt, in a fresh sandbox.
  - [Saved cases](/docs/evals/saved-cases): Save a real turn from a thread as a regression case on disk: its history, the model's replies, tool results and what must happen.
  - [Scripted model & fake sandbox](/docs/evals/testing): Test agents deterministically, offline and with no API keys, using a scripted model and an in-memory sandbox.
- API Reference: Functions and types
  - [API reference](/docs/reference/overview): Every public function and type, in TypeScript and Python.

  - **Functions**
  - [agent](/docs/reference/functions/agent): Describe an agent.
  - [tool](/docs/reference/functions/tool): Define an app tool.
  - [extension](/docs/reference/functions/extension): The one extension primitive: tools, trusted instructions and hooks.
  - [mcp](/docs/reference/functions/mcp): An MCP server in one line.
  - [secret](/docs/reference/functions/secret): A reference to a host secret, resolved only in the host.
  - [openThread / open_thread](/docs/reference/functions/openThread): A handle for inspection and control, independent of any in-memory run.
  - [sqlite](/docs/reference/functions/sqlite): The one log and artifact store.
  - [localMemory / local_memory](/docs/reference/functions/localMemory): Built-in memory: SQLite FTS5 in the run's store.
  - [localKnowledge / local_knowledge](/docs/reference/functions/localKnowledge): Built-in knowledge: SQLite FTS5 over admitted versions stored as artifacts.
  - [scriptedModel / scripted_model](/docs/reference/functions/scriptedModel): Test kit.
  - [fakeSandbox / fake_sandbox](/docs/reference/functions/fakeSandbox): Test kit.
  - [host](/docs/reference/functions/host): Bind agents to a store, channels and schedules.

  - **Agents and runs**
  - [Agent](/docs/reference/types/Agent): An agent made by agent(): its name and config, ready to run.
  - [RunResult](/docs/reference/types/RunResult): The HTTP form is host-api RunOutcome (ids instead of handles).
  - [RunStream](/docs/reference/types/RunStream): What agent.stream() returns.
  - [RunContext](/docs/reference/types/RunContext): What your tool or hook receives about the run it is part of.
  - [Tool](/docs/reference/types/Tool): An app tool (tool()).
  - [McpServer](/docs/reference/types/McpServer): An MCP server binding (mcp()).
  - [Extension](/docs/reference/types/Extension): An extension (extension()).
  - [Hooks](/docs/reference/types/Hooks): Every hook is optional, awaited and bounded by hook_timeout_ms.
  - [Skill](/docs/reference/types/Skill): Loaded from the host config store, pinned by hash.
  - [Secret](/docs/reference/types/Secret): A reference to a host secret, made by secret(name).
  - [Store](/docs/reference/types/Store): The SQLite log and artifact store (sqlite()).
  - [ConfigError](/docs/reference/types/ConfigError): Thrown at setup only: TS class ConfigError extends Error, Python exception ConfigError.
  - [ConfigErrorCode](/docs/reference/types/ConfigErrorCode): Why setup failed.

  - **Threads and evals**
  - [Thread](/docs/reference/types/Thread): A thread positioned at one branch.
  - [CaseExpectation](/docs/reference/types/CaseExpectation): Written into case.json.
  - [SavedCase](/docs/reference/types/SavedCase): What saveCase returns: where the case was written and whether it replays anywhere.

  - **Host**
  - [Host](/docs/reference/types/Host): TS: supports await using (Symbol.asyncDispose calls stop).
  - [Schedule](/docs/reference/types/Schedule): A cron schedule on a host: each time it is due, it starts a run of one of the host's agents on a new thread.

  - **Model adapters**
  - [Model](/docs/reference/types/Model): What a model adapter package (anthropic(), openai(), ...) returns.
  - [ModelContext](/docs/reference/types/ModelContext): What the loop hands an adapter for one send or lookup: the fencing pair it re-checks at its real send point, and artifact access, since the rendered request carries refs, not bytes.
  - [ModelInfo](/docs/reference/types/ModelInfo): What a model adapter declares about itself: which model and adapter it is, its parameters and limits, the inputs it accepts, its hosted tools and whether a lost response can be looked up.
  - [ModelRequest](/docs/reference/types/ModelRequest): One send attempt: the client request id and the exact rendered request bytes the adapter encodes for its provider.
  - [ModelResponse](/docs/reference/types/ModelResponse): A complete model response, as a found lookup returns it: the output parts, why it stopped, token usage and the provider's request id.
  - [ModelChunk](/docs/reference/types/ModelChunk): One item of a Model.send stream: a text delta, a finished output part, the final done chunk, or one terminal rejected chunk when the send failed before any content.
  - [LookupResult](/docs/reference/types/LookupResult): An adapter's answer about an operation it may have performed; each answer carries its own finality.
  - [LookupCapability](/docs/reference/types/LookupCapability): Declared per lookup operation and verified by the provider conformance suite.

  - **Sandbox adapters**
  - [Sandbox](/docs/reference/types/Sandbox): What a sandbox adapter (e2b(), daytona(), modal(), fakeSandbox()) returns.
  - [SandboxAuthority](/docs/reference/types/SandboxAuthority): Who may dispatch a provider operation.
  - [SandboxContext](/docs/reference/types/SandboxContext): What the runtime hands a sandbox adapter for every provider operation: the authority it re-checks at its real dispatch point.
  - [SandboxInfo](/docs/reference/types/SandboxInfo): What a sandbox adapter declares about itself: its provider name, whether it enforces egress, what its snapshots capture, browser and desktop support, which lost operations it can look up, and whether terminate can confirm a process is gone.
  - [SandboxSession](/docs/reference/types/SandboxSession): One live sandbox, as an adapter returns it from create, restore or attach: run commands, move files, snapshot and close it.
  - [ExecOutput](/docs/reference/types/ExecOutput): What an adapter's exec returns: the complete output as byte streams, never truncated or buffered whole.
  - [ExecResult](/docs/reference/types/ExecResult): What tools and the log see.

  - **Memory and knowledge adapters**
  - [MemoryProvider](/docs/reference/types/MemoryProvider): Long-term memory storage behind save_memory, search_memory and forget_memory.
  - [KnowledgeProvider](/docs/reference/types/KnowledgeProvider): Knowledge-base storage behind search_knowledge.
  - [Scope](/docs/reference/types/Scope): From host config and the verified principal.
  - [Binding](/docs/reference/types/Binding): Issued by the host; the provider stores it opaquely.
  - [MemoryRecord](/docs/reference/types/MemoryRecord): One fact to store, as the framework hands it to MemoryProvider.remember: the text, where it came from, the log events behind it, and the host-issued binding.
  - [RecordRef](/docs/reference/types/RecordRef): What MemoryProvider.remember returns: the stored record's id and version.
  - [MemoryHit](/docs/reference/types/MemoryHit): One recalled memory, as a provider returns it from recall.
  - [KnowledgeSource](/docs/reference/types/KnowledgeSource): One document to admit into a knowledge base, as the framework hands it to KnowledgeProvider.ingest.
  - [DocVersion](/docs/reference/types/DocVersion): What KnowledgeProvider.ingest returns: the admitted version of a document and the store revision after the ingest.
  - [KnowledgeHit](/docs/reference/types/KnowledgeHit): One matching passage, as a provider returns it from search.
  - [Doc](/docs/reference/types/Doc): One admitted document version, as KnowledgeProvider.get returns it.
  - [SearchBackend](/docs/reference/types/SearchBackend): A web_search backend adapter (Exa, Brave, Tavily).
  - [SearchHit](/docs/reference/types/SearchHit): One web search result, as a SearchBackend returns it.

  - **Channel adapters**
  - [ChannelAdapter](/docs/reference/types/ChannelAdapter): What a channel adapter (slack(), whatsapp(), github()) returns.
  - [ChannelCapabilities](/docs/reference/types/ChannelCapabilities): What a channel adapter declares about its provider.
  - [RawRequest](/docs/reference/types/RawRequest): A webhook as the host received it.
  - [RawResponse](/docs/reference/types/RawResponse): An HTTP reply an adapter hands back to the host to send.
  - [VerifiedDelivery](/docs/reference/types/VerifiedDelivery): What a provider's signature proves about a webhook: whose workspace or installation sent it.
  - [Inbound](/docs/reference/types/Inbound): One item of a verified webhook batch: a message, an approval decision, a control command, or something to ignore.
  - [DeliveryOutcome](/docs/reference/types/DeliveryOutcome): What perform knows about a send: sent, or a delivery error that says whether anything may have reached the provider.
- HTTP API: Host routes
  - [HTTP API](/docs/http-api): The host's typed JSON API, generated from its OpenAPI file.
  - [Start a run](/docs/http-api/start-run)
  - [Stream a run's events](/docs/http-api/subscribe-run)
  - [Get a thread's timeline](/docs/http-api/get-timeline)
  - [List branches](/docs/http-api/list-branches)
  - [List fork points](/docs/http-api/list-fork-points)
  - [Fork a thread](/docs/http-api/fork)
  - [List pending approvals](/docs/http-api/list-approvals)
  - [Approve or deny](/docs/http-api/decide-approval)
  - [Resolve a parked action](/docs/http-api/resolve-parked)
  - [Cancel a thread](/docs/http-api/cancel)
  - [Change the model](/docs/http-api/set-model)
  - [Change the permission mode](/docs/http-api/set-mode)
  - [Channel verification challenge](/docs/http-api/channel-challenge)
  - [Channel webhook](/docs/http-api/channel-webhook)