Chat API — agent-powered streaming conversations

The Chat API gives you a streaming conversational agent that can take action on your behalf — running market research, expanding build nodes, kicking off Forge coding runs, and calling any connected integration like Linear, Slack, or GitHub. Every message you send streams back NDJSON events as the agent thinks, calls tools, and composes its reply. Sessions are persisted automatically so you can resume a conversation at any time.

Start a chat

POST /chat is the primary endpoint. It accepts your message history and streams back a sequence of NDJSON events until the agent finishes its reply.

This endpoint uses Stream auth. Your session token’s TTL is extended on connect so long-running agentic turns don’t expire mid-stream.

Rate limit: 20 requests per minute.

messages

object[]

required

Full conversation history. Each message has a role (user or assistant) and content string (max 50,000 characters per message).

idea

string

required

The product idea the conversation is focused on. Max 50,000 characters. The agent uses this as persistent context for all tool calls.

project_id

string

UUID of an existing project to scope tool calls (research, build graphs) to. When set, the server prepends the project’s one-line digest to the agent’s system prompt so the agent knows what’s been happening in that project. Pass null to work without a project — the agent falls back to the user’s most recently active project (within the last 24 hours) when one is needed.

session_id

string

ID of an existing session to continue. Pass null to start a new session — the server mints one and sends it back in the stream_start event.

has_research

boolean

default:"false"

Tell the agent whether completed research exists for this project. Helps it decide whether to call run_research or query_research first.

has_build_tree

boolean

default:"false"

Tell the agent whether a build graph exists. Helps it choose between expand_build_node and other tools.

research_context

string

Pre-loaded research text to inject into the agent’s context window. Max 100,000 characters. Useful when you want to avoid a query_research round-trip.

curl -X POST https://api.manticscore.com/chat \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "What are the biggest risks in this market?"}],
    "idea": "AI expense tracker for freelancers",
    "project_id": "9f4e2a1b-...",
    "has_research": true
  }'

Stream events

The response is Content-Type: application/x-ndjson. Each line is a complete JSON object: {"v": 1, "event": "<type>", "data": {...}}.

Parse on the event field and silently ignore any event types you don’t recognize. New events may be added without a version bump.

Event	Data	Description
`stream_start`	`{"request_id": "string", "session_id": "string"}`	First event. Save `session_id` to resume this conversation later.
`agent_turn`	`{"turn": int, "stop_reason": "tool_use\|end_turn"}`	The agent is starting a new turn. `tool_use` means it’s calling a tool next.
`thinking_delta`	`{"turn": int, "text": "string"}`	Optional extended reasoning text. Append to a hidden reasoning buffer.
`tool_call`	`{"tool_call_id": "string", "tool_name": "string", "args": {"turn": int}}`	The agent invoked a tool. Show an activity indicator with the tool name.
`tool_progress`	`{"turn": int, "tool": "string", "status": "in_progress\|completed\|error", "duration_ms?": int, "error?": "string", "job_id?": "string"}`	Live status update for the running tool. `job_id` is set when a pipeline was started.
`message_delta`	`{"text": "string"}`	Append this text to the current message bubble.
`card.signin_required`	`{"toolkit": "string", "...": "..."}`	The agent attempted an action that needs a Composio sign-in. Render the inline card so the user can authorize without leaving the chat.
`card.oauth_required`	`{"toolkit": "string", "auth_url": "string", "...": "..."}`	The user must complete an OAuth flow before the agent can continue. The card carries an `auth_url` to open in a browser.
`card.agent_spawned`	`{"agent_id": "string", "goal": "string", "schedule_cron": "string", "next_run_at": "string", "actions": [{"label": "string", "tool": "string", "args": {}}]}`	Emitted after `spawn_agent` succeeds. Render the agent card with the supplied quick actions (e.g. Pause, Run now) wired back to `update_agent`.
`conversation_summary`	`{"summary": "string", "tools_used": ["string"], "turns": int, "session_id": "string"}`	Emitted at the end of every response. Use to update the session card.
`error`	`{"message": "string"}`	An unrecoverable error occurred. Close the stream and surface the message.
`done`	`{}`	Stream is finished.

card.* events are persisted alongside other session events, so they replay correctly when you re-open a session via GET /chat/sessions/{id}.

Built-in agent tools

The agent has access to the following tools on every request. Composio tools (GitHub, Linear, Slack, Notion, etc.) are discovered at runtime based on the user’s connected accounts — the agent can invoke any of 1000+ tools without a hardcoded list.

Tool	What it does
`query_research`	Searches existing completed research for the current project.
`run_research`	Starts a new market research pipeline and streams its progress.
`expand_build_node`	Expands a collapsed build graph node into child tasks.
`assess_node_risk`	Runs an LLM risk and effort assessment on a build node.
`deep_dive_feature`	Runs a single-feature deep research job.
`deep_research_features`	Runs deep research across multiple features simultaneously.
`start_implementation`	Kicks off a Forge agentic coding run to produce a GitHub PR.
`spawn_agent`	Deploys a long-running autonomous agent that runs the user’s goal on a schedule (cron) or in response to events (`on:research_completed`, `on:signal_detected`, `on:forge_pr_created`, `on:build_graph_completed`, `on:brief_ready`). The agent re-reads its goal each run and discovers Composio tools at runtime. Emits a `card.agent_spawned` event the iOS client renders inline. If a required toolkit isn’t connected yet, emits a `card.oauth_required` instead.
`update_agent`	Edits an existing spawned agent — change `goal`, `schedule`, `toolkits`, or pause/resume by toggling `enabled`. Accepts either a structured `patch` object or a plain-language `natural_language` instruction the server re-prompts to derive the diff.
`query_knowledge`	Looks up everything the system knows about a canonical entity. Pass `canonical_kind` (`company`, `feature`, or `blueprint`) and `canonical_id` to receive the entity record plus every artifact (research, brief, signal, etc.) that references it. Use this when the user asks for cross-research context like “what do you know about Stripe”.

When run_research, expand_build_node, or start_implementation run, the server bridges their pipeline stage events as tool_progress events — you see the full real-time progress inline in the chat stream.

run_research from chat is dispatched through the same background workflow runner as POST /research, so chat-initiated research uses identical pipeline behavior and capacity to direct API calls. The agent waits up to 180 seconds for the pipeline to emit done. If the pipeline hasn’t completed in that window the tool returns {"status": "timeout", ...} to the model — the underlying research job continues running in the background and can be polled or streamed via the standard research endpoints.

deep_research_features is fire-and-forget: the agent dispatches the job through the same background workflow runner as POST /feature-research and returns immediately with {"job_id": "...", "status": "queued"}. Subscribe to GET /feature-research/{job_id}/events to stream progress, or poll GET /feature-research/{job_id}/status.

Sessions

List sessions

limit

number

default:"20"

Maximum sessions to return. Range: 1–100.

offset

number

default:"0"

Pagination offset.

curl "https://api.manticscore.com/chat/sessions?limit=20" \
  -H "Authorization: Bearer <token>"

200 response

[
  {
    "id": "s9f4e2a1b-...",
    "project_id": "9f4e2a1b-...",
    "idea": "AI expense tracker for freelancers",
    "status": "completed",
    "summary": "Discussed market risks and created a Linear issue for the auth module.",
    "message_count": 8,
    "tool_calls_count": 3,
    "created_at": "2026-04-18T14:00:00Z",
    "updated_at": "2026-04-18T14:12:00Z"
  }
]

Get a session

Returns the session object plus its full persisted event log, which you can use to replay or display conversation history.

curl "https://api.manticscore.com/chat/sessions/s9f4e2a1b-..." \
  -H "Authorization: Bearer <token>"

200 response

{
  "id": "s9f4e2a1b-...",
  "project_id": "9f4e2a1b-...",
  "idea": "AI expense tracker for freelancers",
  "status": "completed",
  "summary": "Discussed market risks and created a Linear issue.",
  "message_count": 8,
  "tool_calls_count": 3,
  "events": [
    {"seq": 1, "event_type": "stream_start", "data": {}, "created_at": "2026-04-18T14:00:00Z"},
    {"seq": 2, "event_type": "message_delta", "data": {"text": "The biggest risks are..."}, "created_at": "2026-04-18T14:00:02Z"}
  ]
}

Delete a session

Permanently deletes the session and its event log.

curl -X DELETE "https://api.manticscore.com/chat/sessions/s9f4e2a1b-..." \
  -H "Authorization: Bearer <token>"

Returns 204 No Content on success.

Memory

The agent stores durable memories across sessions — things like preferred naming conventions, team context, and past decisions. Memories are scoped to a project when project_id is provided.

List memories

project_id

string

Scope memories to a specific project UUID. Omit to list all memories.

limit

number

default:"20"

Maximum memories to return. Range: 1–100.

curl "https://api.manticscore.com/chat/memory?project_id=9f4e2a1b-..." \
  -H "Authorization: Bearer <token>"

200 response

{
  "memories": [
    "User prefers Linear over Jira for issue tracking.",
    "Target user is a freelance designer, not a developer."
  ],
  "count": 2
}

Clear memories

Removes all memories for the given scope. Omit project_id to clear all memories across every project.

This action is irreversible. The agent will lose context it has accumulated over past conversations.

curl -X DELETE "https://api.manticscore.com/chat/memory?project_id=9f4e2a1b-..." \
  -H "Authorization: Bearer <token>"

200 response

{"cleared": true}

Overview

Projects & Research

Build & Forge

Intelligence & Chat

Integrations & Ship

Chat API — agent-powered streaming conversations

Start a chat

Stream events

Built-in agent tools

Sessions

List sessions

Get a session

Delete a session

Memory

List memories

Clear memories

Overview

Projects & Research

Build & Forge

Intelligence & Chat

Integrations & Ship

Documentation Index

​Start a chat

​Stream events

​Built-in agent tools

​Sessions

​List sessions

​Get a session

​Delete a session

​Memory

​List memories

​Clear memories

Start a chat

Stream events

Built-in agent tools

Sessions

List sessions

Get a session

Delete a session

Memory

List memories

Clear memories