Shared State
Create a two-way connection between your UI and agent state.
What is shared state?
Shared state connects your UI with the agent's run. With shared state you can:
- Display the agent's current progress and intermediate results
- Update the agent's state through UI interactions
- React to state changes in real-time across your application
When should I use this?
Use shared state when you want the agent and the user to collaborate through the same application state. The agent's outputs are reflected in the UI, and user updates in the UI are reflected in the agent's execution.
Building stateful agents?Persistent threads ship with CopilotKit Intelligence on the free Developer tier.Get CopilotKit Intelligence freeReading agent state
Put shared state in the system prompt and tools
The demo exposes frontend preferences to Claude and gives the agent a tool for writing notes back into CopilotKit state. Keep the state contract small, typed, and mirrored by the UI components that read the same values.
SYSTEM_PROMPT = dedent("""
You are a helpful, concise assistant.
The user's preferences are supplied via shared state and added at the
start of every turn — always respect them. Address the user by name
when known, match the requested tone, and respond in the requested
language.
When the user asks you to "remember" something, or you observe
something worth surfacing in the UI's notes panel, call the
``set_notes`` tool with the FULL updated list of short notes
(existing notes + new ones, not a diff). Keep each note short
(< 120 characters). After updating notes, briefly acknowledge what
you remembered.
""").strip()
SET_NOTES_TOOL: dict[str, Any] = {
"name": "set_notes",
"description": (
"Replace the notes array in shared state with the FULL updated "
"list. Always include every existing note plus any new ones, "
"not a diff. Keep each note short (< 120 chars)."
),
"input_schema": {
"type": "object",
"properties": {
"notes": {
"type": "array",
"items": {"type": "string"},
"description": "Full list of short note strings to persist.",
},
},
"required": ["notes"],
},
}
def build_preferences_block(prefs: dict[str, Any] | None) -> str | None:
"""Render the user-supplied preferences as an injectable prompt block.
Returns ``None`` when no recognised keys are present so the system
prompt is left untouched.
"""
if not isinstance(prefs, dict) or not prefs:
return None
lines = ["The user has shared these preferences with you:"]
if prefs.get("name"):
lines.append(f"- Name: {prefs['name']}")
if prefs.get("tone"):
lines.append(f"- Preferred tone: {prefs['tone']}")
if prefs.get("language"):
lines.append(f"- Preferred language: {prefs['language']}")
interests = prefs.get("interests") or []
if isinstance(interests, list) and interests:
lines.append(f"- Interests: {', '.join(str(i) for i in interests)}")
if len(lines) == 1:
# No recognised fields — don't emit a header with no content.
return None
lines.append(
"Tailor every response to these preferences. Address the user "
"by name when appropriate."
)
return "\n".join(lines)
def _state_dict(state: dict[str, Any]) -> dict[str, Any]:
"""Coerce the AG-UI raw state envelope into the slots we care about."""
return {
"preferences": state.get("preferences") or {},
"notes": list(state.get("notes") or []),
}
Register set_notes with Claude
This dedicated demo uses the Anthropic Messages API directly. Pass the
schema to client.messages.stream so Claude can call set_notes.
stream_request = client.messages.stream(
model=normalize_claude_model(
os.getenv("ANTHROPIC_MODEL", DEFAULT_ANTHROPIC_MODEL)
),
max_tokens=2048,
system=system,
messages=messages,
tools=[SET_NOTES_TOOL],
)Write the tool result back to shared state
When Claude calls set_notes, replace the notes array, emit a
StateSnapshotEvent for the UI, and return a ToolCallResultEvent to
Claude before the loop continues.
# Execute set_notes by mutating shared state and emitting a
# StateSnapshotEvent so the UI re-renders the agent-authored
# notes. This is the agent-side half of the WRITE direction.
tool_results: list[dict[str, Any]] = []
for tc in tool_calls:
if tc["name"] == "set_notes":
notes = tc["input"].get("notes") or []
if isinstance(notes, list):
state["notes"] = [str(n) for n in notes]
result_text = json.dumps({"status": "ok", "count": len(state["notes"])})
yield encoder.encode(
StateSnapshotEvent(type=EventType.STATE_SNAPSHOT, snapshot=state)
)
else:
result_text = json.dumps({"error": f"unknown tool {tc['name']}"})
yield encoder.encode(
ToolCallResultEvent(
type=EventType.TOOL_CALL_RESULT,
tool_call_id=tc["id"],
message_id=f"{msg_id}-tool-result-{tc['id']}",
content=result_text,
)
)
tool_results.append(
{
"type": "tool_result",
"tool_use_id": tc["id"],
"content": result_text,
}
)
messages.append({"role": "user", "content": tool_results})Serve the shared-state loop
Mount the dedicated runner on /shared-state-read-write and return its
AG-UI events as a streaming response.
@app.post("/shared-state-read-write")
async def shared_state_read_write_endpoint(request: Request) -> StreamingResponse:
"""Bidirectional shared state demo — UI writes preferences, agent writes notes.
Uses its own streaming loop (not the shared sales-assistant
``run_agent``) because the state schema, tools, and prompt-injection
middleware are all demo-specific.
"""
body = await request.json()
input_data = RunAgentInput(**body)
async def event_stream() -> AsyncIterator[str]:
async for chunk in run_shared_state_read_write_agent(input_data):
yield chunk
return StreamingResponse(
event_stream(),
media_type="text/event-stream",
headers={
"Cache-Control": "no-cache",
"X-Accel-Buffering": "no",
},
)
This Python route intentionally uses its own Messages API loop. It does not
use ClaudeAgentAdapter or an MCP server.
Subscribe a component to the agent's state with useAgent. Any time the agent mutates its state, for example via a tool call, the composable updates its reactive agent ref and your UI re-renders with the new values.
<script setup lang="ts">
import { computed } from "vue";
import { useAgent, UseAgentUpdate } from "@copilotkit/vue/v2";
type WorkspaceState = {
notes?: string[];
preferences?: { tone?: string };
};
const { agent } = useAgent({
agentId: "workspace",
updates: [UseAgentUpdate.OnStateChanged],
});
const state = computed(
() => (agent.value?.state ?? {}) as WorkspaceState,
);
</script>
<template>
<ul>
<li v-for="note in state.notes ?? []" :key="note">{{ note }}</li>
</ul>
</template>The returned agent.value.state is a plain object. Read it like any other reactive application data and render the parts you care about: agent-written notes, structured outputs, progress indicators, or anything else the agent has put there.
Writing agent state
The same agent object exposes setState. Calling it from a UI event handler pushes the new value into shared state, and the agent reads it back on its next turn. The UI's writes visibly steer the model.
<script setup lang="ts">
function setTone(tone: string) {
const current = agent.value;
if (!current) return;
const previous = (current.state ?? {}) as WorkspaceState;
current.setState({
...previous,
preferences: { ...previous.preferences, tone },
});
}
</script>
<template>
<button @click="setTone('concise')">Prefer concise answers</button>
</template>The UI can also send the agent new inputs, such as preferences, selections, or partial work, without the chat.
Rendering shared state in the UI
Because agent.state is plain data exposed through a Vue ref, the UI layer is ordinary Vue. The example above wires the agent's outputs into a list and feeds user edits back through setState.
Nothing about this is chat-specific: useAgent works in any component under CopilotKitProvider, so you can render shared state in your main view or canvas, not just inside the chat panel. See Render agent state in your app for the full main-view pattern.
<script setup lang="ts">
import { CopilotChat, CopilotKitProvider } from "@copilotkit/vue/v2";
import SharedWorkspace from "./components/SharedWorkspace.vue";
</script>
<template>
<CopilotKitProvider runtime-url="/api/copilotkit">
<SharedWorkspace />
<CopilotChat agent-id="workspace" />
</CopilotKitProvider>
</template>Streaming partial state updates
By default, shared state only updates between backend checkpoints, so a long-running tool call appears as one big burst at the end. State streaming copies a tool argument into a state key while the model writes it, so the UI can watch the answer assemble token-by-token. The sample below is the LangGraph (Python) version.
import json
from collections.abc import AsyncIterator
from typing import Any
from ag_ui.core import EventType, StateSnapshotEvent
from ag_ui.encoder import EventEncoder
from pydantic import BaseModel
class AgentState(BaseModel):
document: str = ""
WRITE_DOCUMENT_TOOL_SCHEMA: dict[str, Any] = {
"name": "write_document",
"description": "Write a document into shared agent state.",
"input_schema": {
"type": "object",
"properties": {
"document": {
"type": "string",
"description": "The full document text to render in shared state.",
},
},
"required": ["document"],
},
}
def _partial_json_string_property(source: str, key: str) -> str | None:
"""Return the current value of a streamed JSON string property."""
marker = json.dumps(key)
key_pos = source.find(marker)
if key_pos < 0:
return None
colon_pos = source.find(":", key_pos)
if colon_pos < 0:
# The colon hasn't streamed yet (e.g. source is `{"document`); bail out
# rather than letting find(":") return -1 and matching the key's own
# opening quote, which would report the key name as the value.
return None
value_start = source.find('"', colon_pos + 1)
if value_start < 0:
return None
raw_chars: list[str] = []
escaped = False
for char in source[value_start + 1 :]:
if escaped:
raw_chars.append("\\" + char)
escaped = False
elif char == "\\":
escaped = True
elif char == '"':
break
else:
raw_chars.append(char)
try:
return json.loads(f'"{"".join(raw_chars)}"')
except json.JSONDecodeError:
return None
async def stream_document_state(
anthropic_stream: AsyncIterator[Any],
state: AgentState,
) -> AsyncIterator[str]:
encoder = EventEncoder()
current_tool_name: str | None = None
current_tool_args = ""
last_streamed_document = state.document
async for event in anthropic_stream:
if (
event.type == "content_block_start"
and event.content_block.type == "tool_use"
):
current_tool_name = event.content_block.name
current_tool_args = ""
continue
if (
event.type != "content_block_delta"
or event.delta.type != "input_json_delta"
):
continue
current_tool_args += event.delta.partial_json
if current_tool_name != "write_document":
continue
streamed_document = _partial_json_string_property(
current_tool_args,
"document",
)
if streamed_document is None or streamed_document == last_streamed_document:
continue
state.document = streamed_document
last_streamed_document = streamed_document
yield encoder.encode(
StateSnapshotEvent(
type=EventType.STATE_SNAPSHOT,
snapshot=state.model_dump(),
)
)
See State streaming for the full walkthrough, including the corresponding useAgent subscription on the frontend.
Read-only context
When a value is UI-owned and the agent should read it but never write it back, such as current user, selected record, or scroll position, reach for useAgentContext instead of full shared state. It publishes values as a one-way UI-to-agent channel that auto-unregisters on unmount.
useAgentContext is a separate read-only capability; it does not replace the bidirectional agent.state / agent.setState(...) contract above.
See Agent read-only context for the full pattern.
How is this guide?