LangChain Reference home pageLangChain ReferenceLangChain Reference
  • GitHub
  • Main Docs
Deep Agents
LangChain
LangGraph
Integrations
LangSmith
LangGraph
  • Web
  • Channels
  • Pregel
  • Prebuilt
  • Remote
React SDK
Vue SDK
Svelte SDK
Angular SDK
LangGraph SDK
  • Ui
  • Client
  • Auth
  • React
  • Logging
  • React Ui
  • Utils
  • Server
LangGraph Checkpoint
LangGraph Checkpoint MongoDB
LangGraph Checkpoint Postgres
  • Store
LangGraph Checkpoint Redis
  • Shallow
  • Store
LangGraph Checkpoint SQLite
LangGraph Checkpoint Validation
  • Cli
LangGraph API
LangGraph CLI
LangGraph CUA
  • Utils
LangGraph Supervisor
LangGraph Swarm
⌘I

LangChain Assistant

Ask a question to get started

Enter to send•Shift+Enter new line

Menu

LangGraph
WebChannelsPregelPrebuiltRemote
React SDK
Vue SDK
Svelte SDK
Angular SDK
LangGraph SDK
UiClientAuthReactLoggingReact UiUtilsServer
LangGraph Checkpoint
LangGraph Checkpoint MongoDB
LangGraph Checkpoint Postgres
Store
LangGraph Checkpoint Redis
ShallowStore
LangGraph Checkpoint SQLite
LangGraph Checkpoint Validation
Cli
LangGraph API
LangGraph CLI
LangGraph CUA
Utils
LangGraph Supervisor
LangGraph Swarm
Language
Theme
JavaScript@langchain/svelte

@langchain/svelte

Description

@langchain/svelte

Svelte SDK for building AI-powered applications with Deep Agents, LangChain and LangGraph. It provides a useStream function that manages streaming, state, branching, and interrupts with a Svelte 5 runes-compatible reactive API.

Installation

npm install @langchain/svelte @langchain/core

Peer dependencies: svelte (^5.0.0), @langchain/core (^1.0.1)

Quick Start

<script lang="ts">
  import { useStream } from "@langchain/svelte";

  const stream = useStream({
    assistantId: "agent",
    apiUrl: "http://localhost:2024",
  });
</script>

<div>
  {#each stream.messages as msg, i (msg.id ?? i)}
    <div>{msg.content}</div>
  {/each}

  <button
    disabled={stream.isLoading}
    onclick={() =>
      void stream.submit({ messages: [{ type: "human", content: "Hello!" }] })}
  >
    Send
  </button>
</div>

All reactive properties (messages, isLoading, values, etc.) are accessed directly on the returned object — no $ prefix needed. Avoid destructuring reactive properties; use stream.messages instead of const { messages } = stream to keep reactivity intact.

useStream Options

Option Type Description
assistantId string Required. The assistant/graph ID to stream from.
apiUrl string Base URL of the LangGraph API.
client Client Pre-configured Client instance (alternative to apiUrl).
messagesKey string State key containing messages. Defaults to "messages".
initialValues StateType Initial state values before any stream data arrives.
fetchStateHistory boolean \| { limit: number } Fetch thread history on stream completion. Enables branching.
throttle boolean \| number Throttle state updates for performance.
onFinish (state, error?) => void Called when the stream completes.
onError (error, state?) => void Called on stream errors.
onThreadId (threadId) => void Called when a new thread is created.
onUpdateEvent (event) => void Receive update events from the stream.
onCustomEvent (event) => void Receive custom events from the stream.
onStop () => void Called when the stream is stopped by the user.

Return Values

All reactive properties are exposed as getters on the returned object. They update automatically and can be read directly in Svelte 5 templates without the $ prefix.

Property Type Description
values StateType Current graph state.
messages BaseMessage[] Messages from the current state.
isLoading boolean Whether a stream is currently active.
error unknown The most recent error, if any.
interrupt Interrupt \| undefined Current interrupt requiring user input.
branch string Active branch identifier.
submit(values, options?) function Submit new input to the graph. When called while a stream is active, the run is created on the server with multitaskStrategy: "enqueue" and queued automatically.
stop() function Cancel the active stream.
setBranch(branch) function Switch to a different conversation branch.
getMessagesMetadata(msg, index?) function Get branching and checkpoint metadata for a message.
switchThread(id) (id: string \| null) => void Switch to a different thread. Pass null to start a new thread on next submit.
queue.entries ReadonlyArray<QueueEntry> Pending server-side runs. Each entry has id (server run ID), values, options, and createdAt.
queue.size number Number of pending runs on the server.
queue.cancel(id) (id: string) => Promise<boolean> Cancel a pending run on the server by its run ID.
queue.clear() () => Promise<void> Cancel all pending runs on the server.

Type Safety

Provide your state type as a generic parameter:

<script lang="ts">
  import { useStream } from "@langchain/svelte";
  import type { BaseMessage } from "langchain";

  interface MyState {
    messages: BaseMessage[];
    context?: string;
  }

  const stream = useStream<MyState>({
    assistantId: "my-graph",
    apiUrl: "http://localhost:2024",
  });
</script>

Typed Interrupts

<script lang="ts">
  import { useStream } from "@langchain/svelte";
  import type { BaseMessage } from "langchain";

  const stream = useStream<
    { messages: BaseMessage[] },
    { InterruptType: { question: string } }
  >({
    assistantId: "my-graph",
    apiUrl: "http://localhost:2024",
  });
</script>

Handling Interrupts

<script lang="ts">
  import { useStream } from "@langchain/svelte";
  import type { BaseMessage } from "langchain";

  const stream = useStream<
    { messages: BaseMessage[] },
    { InterruptType: { question: string } }
  >({
    assistantId: "agent",
    apiUrl: "http://localhost:2024",
  });
</script>

<div>
  {#each stream.messages as msg, i (msg.id ?? i)}
    <div>{msg.content}</div>
  {/each}

  {#if stream.interrupt}
    <div>
      <p>{stream.interrupt.value.question}</p>
      <button onclick={() => void stream.submit(null, { command: { resume: "Approved" } })}>
        Approve
      </button>
    </div>
  {/if}

  <button
    onclick={() =>
      void stream.submit({ messages: [{ type: "human", content: "Hello" }] })}
  >
    Send
  </button>
</div>

Branching

Enable conversation branching with fetchStateHistory: true:

<script lang="ts">
  import { useStream } from "@langchain/svelte";

  const stream = useStream({
    assistantId: "agent",
    apiUrl: "http://localhost:2024",
    fetchStateHistory: true,
  });
</script>

<div>
  {#each stream.messages as msg, i (msg.id ?? i)}
    {@const metadata = stream.getMessagesMetadata(msg, i)}
    {@const branchOptions = metadata?.branchOptions}
    {@const currentBranch = metadata?.branch}

    <div>
      <p>{msg.content}</p>

      {#if branchOptions && currentBranch}
        <button onclick={() => {
          const prev = branchOptions[branchOptions.indexOf(currentBranch) - 1];
          if (prev) stream.setBranch(prev);
        }}>
          Previous
        </button>
        <span>
          {branchOptions.indexOf(currentBranch) + 1} / {branchOptions.length}
        </span>
        <button onclick={() => {
          const next = branchOptions[branchOptions.indexOf(currentBranch) + 1];
          if (next) stream.setBranch(next);
        }}>
          Next
        </button>
      {/if}
    </div>
  {/each}

  <button
    onclick={() =>
      void stream.submit({ messages: [{ type: "human", content: "Hello" }] })}
  >
    Send
  </button>
</div>

Server-Side Queuing

When submit() is called while a stream is already active, the SDK automatically creates the run on the server with multitaskStrategy: "enqueue". The pending runs are tracked in queue and processed in order as each finishes:

<script lang="ts">
  import { useStream } from "@langchain/svelte";

  const stream = useStream({
    assistantId: "agent",
    apiUrl: "http://localhost:2024",
  });
</script>

<div>
  {#each stream.messages as msg, i (msg.id ?? i)}
    <div>{msg.content}</div>
  {/each}

  {#if stream.queue.size > 0}
    <div>
      <p>{stream.queue.size} message(s) queued</p>
      <button onclick={() => void stream.queue.clear()}>Clear Queue</button>
    </div>
  {/if}

  <button
    disabled={stream.isLoading}
    onclick={() =>
      void stream.submit({ messages: [{ type: "human", content: "Hello!" }] })}
  >
    Send
  </button>
  <button onclick={() => stream.switchThread(null)}>New Thread</button>
</div>

Switching threads via switchThread() cancels all pending runs and clears the queue.

Stream Context

Use setStreamContext and getStreamContext to share a single useStream instance across a component tree without prop drilling. This uses Svelte's built-in setContext / getContext under the hood.

Setting context in a parent

Call setStreamContext during component initialisation to provide the stream to all descendants:

<script lang="ts">
  import { useStream, setStreamContext } from "@langchain/svelte";
  import ChatMessages from "./ChatMessages.svelte";
  import ChatInput from "./ChatInput.svelte";

  const stream = useStream({
    assistantId: "agent",
    apiUrl: "http://localhost:2024",
  });

  setStreamContext(stream);
</script>

<ChatMessages />
<ChatInput />

Consuming context in a child

Call getStreamContext in any descendant to retrieve the stream. The returned value has the same shape and types as the useStream return value:

<script lang="ts">
  import { getStreamContext } from "@langchain/svelte";

  const stream = getStreamContext();
</script>

{#each stream.messages as msg, i (msg.id ?? i)}
  <div>{msg.content}</div>
{/each}

{#if stream.isLoading}
  <p>Thinking…</p>
{/if}
<script lang="ts">
  import { getStreamContext } from "@langchain/svelte";

  const stream = getStreamContext();

  let input = $state("");
</script>

<form onsubmit={(e) => { e.preventDefault(); void stream.submit({ messages: [{ type: "human", content: input }] }); input = ""; }}>
  <input bind:value={input} />
  <button type="submit">Send</button>
</form>

Type safety

getStreamContext accepts the same generic parameters as useStream so child components can be fully typed:

<script lang="ts">
  import { getStreamContext } from "@langchain/svelte";
  import type { BaseMessage } from "@langchain/core/messages";

  interface MyState {
    messages: BaseMessage[];
    context?: string;
  }

  const stream = getStreamContext<MyState>();
</script>

setStreamContext returns the stream it was given, so you can inline both calls:

<script lang="ts">
  import { useStream, setStreamContext } from "@langchain/svelte";

  const stream = setStreamContext(
    useStream({ assistantId: "agent", apiUrl: "http://localhost:2024" }),
  );
</script>

Note: Both functions must be called during component initialisation (i.e. at the top level of a <script> block), just like Svelte's own setContext / getContext. Calling getStreamContext without a parent setStreamContext throws an error.

Custom Transport

Instead of connecting to a LangGraph API, you can provide your own streaming transport. Pass a transport object instead of assistantId to use a custom backend:

<script lang="ts">
  import { useStream, FetchStreamTransport } from "@langchain/svelte";
  import type { BaseMessage } from "langchain";

  const stream = useStream<{ messages: BaseMessage[] }>({
    transport: new FetchStreamTransport({
      url: "https://my-api.example.com/stream",
    }),
    threadId: null,
    onThreadId: (id) => console.log("Thread created:", id),
  });
</script>

<div>
  {#each stream.messages as msg, i (msg.id ?? i)}
    {@const metadata = stream.getMessagesMetadata(msg, i)}
    <div>
      <p>{msg.content}</p>
      {#if metadata?.streamMetadata}
        <span>Node: {metadata.streamMetadata.langgraph_node}</span>
      {/if}
    </div>
  {/each}

  <p>Current branch: {stream.branch}</p>

  <button
    disabled={stream.isLoading}
    onclick={() =>
      void stream.submit({ messages: [{ type: "human", content: "Hello!" }] })}
  >
    Send
  </button>
</div>

The custom transport interface returns the same properties as the standard useStream function, including getMessagesMetadata, branch, setBranch, switchThread, and all message/interrupt/subagent helpers. When using a custom transport, getMessagesMetadata returns stream metadata sent alongside messages during streaming; branch and setBranch provide local branch state management. onFinish is also supported and receives a synthetic ThreadState built from the final locally streamed values; the run metadata argument is undefined.

Sharing State with provideStream

When multiple components need access to the same stream (a message list, a header, an input bar), use provideStream and getStream to share a single stream instance via Svelte's context API:

<!-- ChatContainer.svelte -->
<script lang="ts">
  import { provideStream } from "@langchain/svelte";
  import ChatHeader from "./ChatHeader.svelte";
  import MessageList from "./MessageList.svelte";
  import MessageInput from "./MessageInput.svelte";

  provideStream({
    assistantId: "agent",
    apiUrl: "http://localhost:2024",
  });
</script>

<ChatHeader />
<MessageList />
<MessageInput />
<!-- MessageList.svelte -->
<script lang="ts">
  import { getStream } from "@langchain/svelte";

  const stream = getStream();
</script>

{#each stream.messages as msg (msg.id)}
  <div>{msg.content}</div>
{/each}
<!-- MessageInput.svelte -->
<script lang="ts">
  import { getStream } from "@langchain/svelte";

  const stream = getStream();
  let input = $state("");

  function send() {
    stream.submit({ messages: [{ type: "human", content: input }] });
    input = "";
  }
</script>

<form onsubmit={send}>
  <textarea bind:value={input}></textarea>
  <button disabled={stream.isLoading} type="submit">Send</button>
</form>
<!-- ChatHeader.svelte -->
<script lang="ts">
  import { getStream } from "@langchain/svelte";

  const stream = getStream();
</script>

<header>
  <h1>Chat</h1>
  {#if stream.isLoading}
    <span>Thinking...</span>
  {/if}
  {#if stream.error}
    <span>Error occurred</span>
  {/if}
</header>

Multiple Agents

Nest provideStream calls for multi-agent scenarios — Svelte's context scoping ensures each subtree gets its own stream:

<!-- ResearchPanel.svelte -->
<script lang="ts">
  import { provideStream } from "@langchain/svelte";
  provideStream({ assistantId: "researcher", apiUrl: "http://localhost:2024" });
</script>

<MessageList />
<MessageInput />

Playground

For complete end-to-end examples with full agentic UIs, visit the LangChain UI Playground.

License

MIT

Classes

Class

FetchStreamTransport

Transport used to stream the thread.

Class

SubagentManager

Manages subagent execution state.

Functions

Function

calculateDepthFromNamespace

Calculates the depth of a subagent based on its namespace.

Function

executeHeadlessTool

Function

extractParentIdFromNamespace

Extracts the parent tool call ID from a namespace.

Function

extractToolCallIdFromNamespace

Extracts the tool call ID from a namespace path.

Function

filterOutHeadlessToolInterrupts

Strip headless-tool interrupts from a user-facing interrupt list.

Function

findHeadlessTool

Function

flushPendingHeadlessToolInterrupts

Execute and resume all newly seen headless-tool interrupts from a values

Function

getStream

Retrieves the shared stream instance from the nearest ancestor that

Function

getStreamContext

Retrieves the useStream instance previously provided by a parent

Function

handleHeadlessToolInterrupt

Function

headlessToolResumeCommand

Function

isHeadlessToolInterrupt

Function

isSubagentNamespace

Checks if a namespace indicates a subagent/subgraph message.

Function

parseHeadlessToolInterruptPayload

Parses a headless-tool interrupt value from the graph. Accepts both

Function

provideStream

Creates a shared useStream instance and makes it available to all

Function

setStreamContext

Provides a useStream return value to all descendant components via

Function

useStream

Interfaces

Types

Interface

AgentTypeConfigLike

Minimal interface matching the structure of AgentTypeConfig from @langchain/langgraph.

Interface

BaseStream

Base stream interface shared by all stream types.

Interface

CompiledSubAgentLike

Minimal interface matching the structure of a CompiledSubAgent from deepagents.

Interface

DeepAgentTypeConfigLike

Minimal interface matching the structure of DeepAgentTypeConfig from deepagents.

Interface

FlushPendingHeadlessToolInterruptsOptions

Interface

HeadlessToolImplementation

Client-side implementation returned by headlessTool.implement(...).

Interface

HeadlessToolInterrupt

Represents a headless tool interrupt payload emitted by LangChain's

Interface

QueueEntry

A single queued submission entry representing a server-side pending run.

Interface

QueueInterface

Reactive interface exposed to framework consumers for observing

Interface

SubagentApi

Subagent API surface parameterised by the subagent interface type.

Interface

SubAgentLike

Minimal interface matching the structure of a SubAgent from deepagents.

Interface

SubagentStreamInterface

Base interface for a single subagent stream.

Interface

SubagentToolCall

Represents a tool call that initiated a subagent.

Interface

ToolEvent

Interface

UseAgentStream

Stream interface for ReactAgent instances created with createAgent.

Interface

UseAgentStreamOptions

Options for configuring an agent stream.

Interface

UseDeepAgentStream

Stream interface for DeepAgent instances created with createDeepAgent.

Interface

UseDeepAgentStreamOptions

Options for configuring a deep agent stream.

Interface

UseStreamOptions

Interface

UseStreamThread

Interface

UseStreamTransport

Transport used to stream the thread.

Type

AnyHeadlessToolImplementation

Type

BaseSubagentState

Base state type for subagents.

Type

ClassSubagentStreamInterface

Subagent stream interface with messages typed as BaseMessage[]

Type

DefaultSubagentStates

Default subagent state map used when no specific subagent types are provided.

Type

DefaultToolCall

Default tool call type when no specific tool definitions are provided.

Type

ExtractAgentConfig

Extract the AgentTypeConfig from an agent-like type.

Type

ExtractDeepAgentConfig

Extract the DeepAgentTypeConfig from a DeepAgent-like type.

Type

ExtractSubAgentMiddleware

Helper type to extract middleware from a SubAgent definition.

Type

GetToolCallsType

Extract the tool call type from a StateType's messages property.

Type

InferAgentToolCalls

Extract tool calls type from an agent's tools.

Type

InferBag

Infer the Bag type from an agent, defaulting to the provided Bag.

Type

InferDeepAgentSubagents

Extract the Subagents array type from a DeepAgent.

Type

InferNodeNames

Infer the node names from a compiled graph.

Type

InferStateType

Infer the state type from an agent, graph, or direct state type.

Type

InferSubagentByName

Helper type to extract a subagent by name from a DeepAgent.

Type

InferSubagentNames

Extract all subagent names as a string union from a DeepAgent.

Type

InferSubagentState

Infer the state type for a specific subagent by extracting and merging

Type

InferSubagentStates

Infer subagent state map from a DeepAgent.

Type

InferToolCalls

Infer tool call types from an agent.

Type

IsAgentLike

Check if a type is agent-like (has ~agentTypes phantom property).

Type

IsDeepAgentLike

Check if a type is a DeepAgent (has ~deepAgentTypes phantom property).

Type

MessageMetadata

Type

OnToolCallback

Type

ResolveStreamInterface

Resolves the appropriate stream interface based on the agent/graph type.

Type

ResolveStreamOptions

Resolves the appropriate options interface based on the agent/graph type.

Type

SubagentStateMap

Create a map of subagent names to their state types.

Type

SubagentStatus

The execution status of a subagent.

Type

SubagentStream

Represents a single subagent stream.

Type

ToolCallFromTool

Infer a tool call type from a single tool.

Type

ToolCallsFromTools

Infer a union of tool call types from an array of tools.

Type

ToolCallState

The lifecycle state of a tool call.

Type

ToolCallWithResult

Type

UseStreamCustomOptions