BrandBrain Flow Orchestrator

Add one node to the stored flow graph. Wire it with connect_flow_nodes, then validate_asset_flow before running.

Upload a local image file (png/jpeg/webp/gif/svg, ≤10 MB) as a session-scoped reference — the backend stores it immutably and the reference node points at its storage key. Use connectToNodeId to feed a generate node, then validate_asset_flow before running.

Apply a batch of graph edits (add/update/remove nodes, connect/disconnect edges) atomically to the stored flowSpec. Preferred over individual calls for multi-step edits; validate afterwards.

Approve the exact manifest revision/hash AND start its run in one durable idempotent call. In live mode this is a paid provider call — confirm the exact scope and cost with the user first.

Record the user's explicit approval of the exact persisted manifest revision and content hash. Never call before the user has seen and approved that manifest; approval alone does not start a run.

Clone a single-generate-node flow once per candidate model (same prompt/style/references), preflight each, and start sibling runs to compare engines. live charges per sibling and needs a total ceiling.

Create a directed edge from one node to another in the stored flow graph; invalid node-kind pairs are rejected.

Create a persisted flow session from a templateId, a free-text goal, or an explicit flowSpec. A goal is patched into the scaffold's brief prompt node — asset-set templates have none, so express intent via update_asset_set_plan instead. Returns the stored session including canvas_url to share with the user.

Remove a directed edge between two nodes in the stored flow graph.

Replace a session's title/description, and optionally REPLACE the entire stored flowSpec (full overwrite — get_asset_flow first). For targeted graph changes prefer add_flow_node, update_flow_node, connect_flow_nodes, or apply_flow_edits.

Estimate what a flow's generate nodes would cost to run, priced from the image-model catalogue (1024² high estimate per image). Pass targetNodeId to scope to the single node a run would execute — one run bills only its target. Free to call — use before a live run to sanity-check spend, or to compare models before set_generate_target.

Fetch a stored session with its full flowSpec graph. Inspect before editing; use list_asset_flows when you only need metadata across flows.

Inspect a stored run's status, per-node traces, artifacts and errors. Use to confirm the real outcome of a finished or failed task before reporting it.

Poll a task started by start_asset_flow_run or start_regenerate_asset_flow_node until completed/failed. Then verify the persisted outcome with get_asset_flow_run — a task report alone is not proof.

Fetch an asset-set plan's manifest revisions and approval events. Use to audit what was proposed and approved before any run.

Fetch a run's durable outputs with metadata, artifact_url (fetchable with this server's auth), open_in_browser link for humans, and a bounded inline preview. The way to view results — never regenerate to fetch outputs.

Report local processor readiness and pinned revisions (split, background removal, vectorisation). Required check before relying on zero-token local stages in asset-set flows.

Get a run's per-node trace: status, duration, resolved prompts, provider request details, artifacts, errors. Use to diagnose a failed or surprising run.

List recorded runs for a session (id, mode, status, timing). Use to find a runId for get_asset_flow_run, get_flow_trace, or get_flow_outputs.

List canonical starter flow templates (id, title, pattern). First step when the user describes a goal but no existing flow; pass the chosen template id to create_asset_flow.

List stored flow canvases (id, title, revision, updated_at, node/edge counts, canvas_url). query and limit filter client-side; archived:true lists read-only retained history. Use to find or resume an existing flow.

List every provider's image models with lifecycle status, reference-image support, and the preflight cost ceiling for one 1024² image. Call before choosing a generate model — the only way to discover OpenRouter model ids. Supports sortByCost and maxPerProvider to surface the cheapest options quickly.

Open BrandBrain sign-in in the user's browser and store a local session shared by MCP clients on this machine. Call when a tool reports not authenticated; verify afterwards with whoami_brandbrain.

Revoke the stored BrandBrain session and remove local credentials for every client sharing it. Use to switch accounts or end access.

Open the flow's canvas editor in the user's default browser and return its URL. Side effect: opens a window on the user's machine — when they only need the link, return canvas_url from list_asset_flows instead.

Open the flow's run gallery (every run's output images with per-image cost and trace) in the user's default browser and return its URL. Opens a window on the user's machine.

Regenerate a single target node in the stored flow session as a task-backed run. Requires MCP task augmentation; clients without it should call start_regenerate_asset_flow_node and poll get_asset_flow_task instead.

Remove a node (and its edges) from the stored flow graph. Run validate_asset_flow afterwards.

Execute the stored flow session in mock or live mode as a task-backed run. Requires MCP task augmentation; clients without it should call start_asset_flow_run and poll get_asset_flow_task instead.

Set the output target on a generate node (surface, aspect ratio, size hint, composition/text policy). Run validate_asset_flow afterwards.

Start a background mock or live run; returns a task to poll with get_asset_flow_task. live spends real provider money — confirm scope and cost with the user first; refused unless BRANDBRAIN_MAX_PROVIDER_SPEND_USD_MICROS is set in the server env. One run executes only the targeted generate node (targetNodeId): downstream generate nodes are paid-call boundaries — run them as separate targeted runs, each billed independently.

Start a background regeneration from a target node; poll it with get_asset_flow_task. Warning: targeting an output node re-runs its paid generate producer — use get_flow_outputs on a finished run to fetch results instead.

Atomically update an asset-set plan node's authoring fields and persist a complete draft manifest proposal; requires expected revisions from the current session. Configure before asking the user to approve.

Update fields on one node in the stored flow graph (prompt text, model, processor options). Run validate_asset_flow afterwards.

Run backend validation on the stored graph. Call after every material edit and before every run — invalid graphs are refused at run time.

Show the stored BrandBrain session (identity and expiry, never tokens). Call first to confirm the signed-in agent identity before editing or running flows.