Server reference

This page is for developers, admins, and anyone building an agent on Continuum. You don’t need any of it to use Continuum from Claude or ChatGPT; Connect your assistant is enough. For the tools themselves, see the tool reference.

Endpoints

Endpoint	Use it for
`https://api.trycontinuum.ai/mcp`	Claude, Claude Code, Cursor, VS Code, Codex, and every other MCP client.
`https://api.trycontinuum.ai/mcp/chatgpt`	ChatGPT. Same tools and sign-in. Long results come back as a one-line summary, with the full data in `structuredContent`, and images aren’t inlined.

The server identifies itself as continuum-mcp.

Transport

Continuum speaks Streamable HTTP.

Request	What it does
`POST`	Sends JSON-RPC messages. Responses are plain JSON.
`GET`	Opens a server-sent events stream for notifications, such as progress and tool-list changes. A comment is sent every 15 seconds to keep it open, and a dropped stream can resume with `Last-Event-Id`.
`DELETE`	Ends the session.

Sessions

The server returns an Mcp-Session-Id on initialize. Send it on every later request.
A request without a session ID that isn’t an initialize gets 400.
A session with no activity and no open stream for 60 minutes is closed. If a request returns 404 with “Session expired or unknown”, send initialize again.
The server’s instructions on initialize include a short guide for the assistant and the starting brand’s published pipelines.
The tool list changes with your access, and the server sends notifications/tools/list_changed when it does.

Authentication

Sign-in is OAuth 2.1. There are no API keys.

Detail	Value
Flow	Authorization code with PKCE (`S256`). Clients are public; there’s no client secret.
Client registration	Dynamic client registration, so clients register themselves.
Discovery	A `401` carries `WWW-Authenticate: Bearer realm="continuum-mcp", resource_metadata="…"`. Metadata is at `/.well-known/oauth-protected-resource/mcp` (or `…/mcp/chatgpt`) and `/.well-known/oauth-authorization-server`.
Token	Sent as `Authorization: Bearer <token>`. Clients refresh it on their own.
Consent	The person signs in to Continuum, picks a brand, and chooses Allow access or Deny.

What a token can do comes from the person’s brands and roles, not from OAuth scopes. Each person sees and revokes their connections under Settings → Connections.

Brand scope

Every call works on one brand.

A session starts on the brand the person last chose in their assistant, otherwise their active brand in the app, otherwise the first brand they can reach.
brands_select switches the brand for the rest of the session. brands_list shows the choices.
To act on another brand for one call, set _meta.brand_id. Access is still checked.
Every result includes _brandContext, with the brand and a link to its Planner.

Roles

Role	Can call
Viewer	Reads: analytics, the brand book, the Library, the calendar, Forge, trends, and competitors. Asking Jaina or the Organic agent (`agent_ask`) needs Operator.
Operator	Reads, plus writes: drafts, publishing, the canvas, pipelines, video edits, ads, and Paid optimization approvals.
Admin, Owner	Everything an Operator can, plus final Forge renders.

A tool whose reads need a connection, such as Meta, appears only when a brand you can reach has that connection.

Confirmations

Publishing, ad and budget changes, live creative swaps, final Forge renders, and most deletions take two calls:

Preview

The first call returns a preview, a confirm_token (tok_…), and its expires_at. Nothing has happened yet. Tools that take a mode input use mode: "preview" for this call.

Confirm

After the person says yes, repeat the call with the same arguments and the confirm_token (and mode: "confirm" where the tool takes a mode).

Detail	Behavior
Lifetime	Set per action and returned as `expires_at`: 10 minutes for ad changes and bulk calendar status changes, 15 minutes for a post approved for publishing, 30 minutes for a `content_reference` remake, 24 hours for a proposed week plan, 5 minutes for the rest.
Single use	Each token works once.
Bound	Tied to the person, the brand, and the exact arguments. Changed arguments return `VALIDATION_ERROR`.
Missing or expired	Returns `CONFIRMATION_REQUIRED`. Preview again.
Admin approval	Activating ads, raising a budget past the brand’s spend limit (by default, more than 30% at once), or any write on a brand with an approval chain returns `APPROVAL_REQUIRED`. An Admin or Owner grants that exact token with `mcp_approval_grant {"token": "…"}` (`mcp_approvals_list` shows what’s waiting); then confirm as usual.

Forge uses its own pair: render_preview or scene_fill returns a confirmationToken, and render_confirm or scene_render redeems it. Forge set_delete is the one deletion with no preview: it deletes immediately, so call it only after an explicit yes. The tool reference lists each tool’s confirmation behavior. Clients that support elicitation may also show a confirmation dialog for ad changes. See You stay in control for the full list of what waits and what happens right away.

Results that ask for something

These come back as normal results, not errors. Each tells the assistant what to do next.

Status	Meaning	Next step
`needs_brand`	No brand is active.	`brands_select`
`needs_integration`	The brand needs a connection, named in `platform`.	`integrations_connect` returns a single-use link that works for 10 minutes.
`needs_reauth`	A connection has lapsed. Reads still work; writes don’t.	Reconnect with the link in `fix`.
`needs_role`	The person’s role is below `required_role`.	Ask an Admin or Owner.
`needs_account`	Several accounts match.	Pick one from `accounts_list`.
`needs_workspace`, `needs_open_canvas`	The canvas has to be open in Continuum.	Open `open_url`, then retry.
`needs_correction`	A question went to the wrong agent.	Ask the `suggested_target`.

Errors

Errors set isError: true. structuredContent holds { ok: false, code, message, details }.

Code	Meaning
`UNAUTHENTICATED`	The token is missing or invalid.
`SESSION_NOT_INITIALIZED`	Send `initialize` first.
`BRAND_NOT_SELECTED`	No brand is active. Tools that are always listed report the same thing as the `needs_brand` status; either way, call `brands_select`.
`PERMISSION_DENIED`, `BRAND_FORBIDDEN`, `ACCOUNT_FORBIDDEN`	The person can’t reach that brand, account, or action.
`VALIDATION_ERROR`	The inputs are wrong, or they changed since a preview.
`CONFIRMATION_REQUIRED`, `APPROVAL_REQUIRED`	See Confirmations.
`INTEGRATION_REQUIRED`, `INTEGRATION_DISCONNECTED`	Connect or reconnect the platform.
`RATE_LIMITED`	Wait `details.retry_after_seconds`. See Limits.
`NOT_FOUND`	The item doesn’t exist, or the brand can’t see it.
`CONFLICT`, `IDEMPOTENCY_CONFLICT`	The item changed, or a repeated request didn’t match the first.
`MEDIA_INVALID`	The file can’t be used.
`VOICE_VIOLATION`	The copy breaks the brand voice.
`META_API_ERROR`, `UPSTREAM_ERROR`	Meta or another service returned an error.
`JOB_PENDING`	The job is still running.
`INTERNAL`	Something went wrong on our side. Contact support if it repeats.

Limits

Limits protect your budget and your connected accounts. RATE_LIMITED includes retry_after_seconds, and per-tool limits also include tool, budget, and window_seconds.

Limit	Current value	Per
Calls, all tools	600 a minute	Person and brand
Canvas runs (`studio_workflow`)	10 an hour	Brand
Trend discovery	10 an hour	Brand
Connection links	10 an hour	Person
Meta ad changes (`paid_manage`)	60 an hour	Brand
Publishing / approvals / deletions (`content_create`)	30 / 60 / 30 an hour	Brand
Calendar changes / week plans (`planner_manage`)	120 / 20 an hour	Brand
Pipeline runs	10 per call	Call

Jobs

Reports, agent answers, and generations run as jobs and return a job_id.

jobs_status takes only job_id and returns pending, running, completed, failed, or cancelled. Leave a few seconds between checks rather than polling in a tight loop. For an agent answer, result.status is completed (it wrote a draft or plan) or needs_followup (text only).
Pass _meta.progressToken on the call to receive notifications/progress.
Pipeline runs report through pipeline_run_status: queued, running, completed, refused, failed, or indeterminate.

Prompts

Prompt	Arguments, all optional
`analytics-weekly-review`	`date_window`
`analytics-top-and-bottom-posts`	`platform`, `metric`, `date_window`
`analytics-top-and-bottom-ads`	`platform`, `kpi`, `date_window`

Activity log

Every tool call is logged with the tool, action, status (ok, error, denied, or rate_limited), duration, brand, and client. Arguments are stored only as a fingerprint, never in full. Each person sees their own calls under Settings → Activity, filtered by status, client, or tool.

Get started

AI assistant (MCP)

For developers and agents

Creative+

Organic+

Performance+

Brand and account

Resources