Tools

The server registers about seventy tools across read, write, and execute scopes. Every tool ships with a structured description (WHAT it does / WHEN to use / WHAT to call NEXT) so the model gets useful guidance without you spelling each call out.

Discovery

The first thing a new chat does is load the orisu-workflows skill from orisu://skills/workflows. The skill teaches the patterns; the tools execute them.

Tool	Scope	Use it when
`search`	per kind	Find an agent, app, model, node, asset, run, or published template by name fragment (scope follows the kind; `node`/`model`/`template` need none).
`who_am_i`	none	Confirm which org the connector is bound to.
`suggest_model`	`agents:read`	Pick a model variant for a `kind` (`image_generation`, etc.) + `use_case`.
`get_effective_ports`	`agents:read`	Resolve the actual port list for a node type given its config (some nodes have dynamic-arity ports). Includes per-port fan-out flags so you can see where lists expand.
`get_run_inputs_schema`	`agents:read`	Discover what `inputs` an agent expects before `trigger_run`.
`get_agent`	`agents:read`	Read an agent’s graph, latest version, and its inputs and outputs as JSON, without rendering the canvas.
`list_agent_versions`	`agents:read`	List an agent’s saved versions (newest first) before `restore_version`.
`get_app`	`apps:read`	Read a published app: the agent behind it, the version it’s pinned to, and its input fields.
`describe`	per kind	Read reference data by `kind`. `node`: one node type’s ports and config fields (leave out `id` for the full catalog). `model`: one model variant’s capabilities, options and pricing (leave out `id` for the catalog). `template`: a published template’s graph to copy. `brand_kit`: one kit’s colors, fonts and voice (leave out `id` to list kits). `integrations`: whether Google Drive or another integration is connected. `credits`: the org’s balance. `playbook`: the workflow skill. `brand_kit` and `integrations` need `agents:read`, `credits` needs `runs:read` or `agents:read`, the rest need none.
`estimate_run_cost`	`agents:read`	Preview credit cost before kicking off a run. Fan-out nodes are priced per item; when the item count is only known at runtime the total is flagged as a lower bound. Also returns `recommended_max_credits`, the value to pass as `max_credits` so the run’s budget check accepts every step.

Building & editing graphs

Tool	Scope	Use it when
`view_agent`	`agents:read`	Show an agent’s current graph. Returns structured nodes/edges AND renders the canvas inline.
`create_agent`	`agents:write`	Build a new agent from a `GraphSpec`. Auto-layouts, validates, saves.
`update_node_config`	`agents:write`	Patch ONE node: `config` is merged (send only changed fields), `value` replaces an input node’s content, `label` renames. Positions preserved. Writes are revision-checked — a concurrent edit fails with `VERSION_CONFLICT` instead of overwriting it. Pass `include_graph: false` to skip the full graph in the response when chaining many edits. The right tool for most edits.
`set_input_values`	`agents:write`	Seed MANY input nodes in one call: a map of node id → value, applied as a single save and version. Unknown or non-input ids are reported per node instead of failing the batch. The fast way to fill in a workflow’s inputs after building it.
`replace_graph`	`agents:write`	Replace an existing agent’s graph with a new `GraphSpec` — for structural changes (add/remove/rewire nodes). Pass `base_version_id` (from `view_agent`’s `latest_version_id`) so concurrent edits fail with `VERSION_CONFLICT` instead of being overwritten.
`validate_graph`	`agents:read`	Validate a spec before saving. Returns typed errors with `next_action` hints.
`preview_node`	`agents:read`	Run one logic node (`prompt_concatenator`, `split_text`, `if_else`, `list_selector`, `text_prompt`, `rename_asset`) in isolation with sample inputs — zero credits, nothing persisted. Model-billed nodes aren’t previewable.
`update_agent`	`agents:write`	Update agent metadata (name, description, etc.).
`clone_agent`	`agents:write`	Duplicate an existing agent.
`delete_agent`	`agents:write`	Soft-delete an agent.
`restore_version`	`agents:write`	Roll back to a saved version.
`create_brand_kit`	`agents:write`	Create a brand kit (name, colors, fonts, logo, voice, guidelines) for brand-aware generation.
`update_brand_kit`	`agents:write`	Patch an existing brand kit — only the fields you pass change.

Running & monitoring

Tool	Scope	Use it when
`run_workflow`	`runs:execute` + `runs:read` (+ `agents:write` to create)	The quickest path. Pass `agent_id` to run an existing agent, or `graph` + `name` to build a new one. It validates the graph, checks the cost, runs it and waits for the result, all in one call. If the graph has errors, nothing is created. `max_credits` defaults to the estimate’s `recommended_max_credits`. Set `confirm_above_credits` to get the estimate back (`COST_CONFIRMATION_REQUIRED`) instead of spending above that amount. A low balance returns `INSUFFICIENT_CREDITS` before anything is created. Returns the same result as `wait_for_run`, plus `agent_id`, `created_agent`, `estimate` and `timed_out`. Shows results inline.
`trigger_run`	`runs:execute`	Run an agent with optional `inputs`. Pre-checks credits (`INSUFFICIENT_CREDITS` fails fast with the estimate). Accepts an optional `max_credits` spend ceiling — the run stops with `RUN_BUDGET_EXCEEDED` before exceeding it. Returns the new `run_id`.
`trigger_node_run`	`runs:execute`	Run an agent up to one node: the target plus everything it depends on, reusing cached unchanged upstream steps. Nothing downstream runs. Takes no input overrides — set values with `update_node_config` first.
`trigger_app`	`apps:execute`	Run a published app with the values its input schema expects. Also accepts `max_credits`.
`wait_for_run`	`runs:read`	Block until a run reaches `completed` / `failed` / `cancelled` / `paused`. Pass `timeout_ms` up to 110000 to bound the wait. Almost always preferable to polling `get_run`. Shows the run’s results inline.
`get_run`	`runs:read`	Read a run’s status, error, and outputs right now, without waiting. Use it to see why a past run failed.
`cancel_run`	`runs:execute`	Cancel an in-progress run.
`list_runs`	`runs:read`	List recent runs, filtered by `agent_id` / `status`.
`submit_review`	`runs:execute`	Approve or reject a run paused at a `human_review` node — the run resumes after.

Tool	Scope	Use it when
`publish_app`	`apps:execute`	Turn an agent into a public app with a defined `inputSchema`.
`share_agent` / `unshare_agent`	`agents:write`	Generate or revoke a read-only share link.
`upload_asset`	`assets:write`	Upload an image / video / audio / file as an Orisu asset that downstream nodes can reference.

Feeds

The feed tools (list_feeds, add_feed_column, run_feed_column, review_feed_cell and the rest) are covered in the Feed guide. One default differs from the grid: a run column added with add_feed_column accepts finished cells automatically (reviewPolicy: "none"), so an unattended batch doesn’t stop to wait for approval. Pass reviewPolicy: "always" when someone should approve each cell, or change it later with update_feed_column. Run columns you add in the grid still default to review.

Inline media

In Claude and ChatGPT, the images, video and audio a run produces appear right in the conversation instead of only as links. The viewer opens for run_workflow, wait_for_run, get_run, trigger_run, trigger_app, preview_asset, list_assets and get_feed_cell_outputs.

After trigger_run or trigger_app, the viewer shows a paper-plane loader and checks on the run every few seconds. Results appear when the run finishes, even before the assistant calls wait_for_run.
Videos and audio play in place. Click a result to see it larger, or full screen when your chat app supports it.
When a run fails, the viewer names the step that failed and shows its error.
The assistant still receives the full JSON with every asset URL, so nothing changes in how it reads results.

The first time the viewer appears, your chat app may ask you to allow it. Chat apps without MCP Apps support get each result as a link instead. For the workflow graph itself, see the in-chat canvas.

Workflow skill

The server’s top-level instructions field directs new agents to run_workflow for build-and-run requests, and to describe({ kind: 'playbook' }) (the same content as the orisu://skills/workflows resource) for a markdown playbook covering:

The mental model (agents, nodes, ports, runs, apps, versions, assets)
7 canonical workflow patterns (text→image, image→video, brand-aware multimodal, branching, fan-out, human-in-the-loop, composer layering)
Build / edit / run / reuse / wait / iterate playbooks
An error-recovery playbook keyed on the typed error.code returned from failed tools
Cost-aware patterns and anti-patterns

It’s served directly off the MCP server (not an external HTTPS URL) so it can’t drift across deploys and stays consistent with the server version you’re talking to.

Resources

The orisu:// reads your assistant should make before designing.

Errors

The error codes, what they mean, and how to recover.

Get started

Reference

Operations

Discovery

Building & editing graphs

Running & monitoring

Feeds

Inline media

Workflow skill

See also

Resources

Errors

Discovery

Building & editing graphs

Running & monitoring

Publishing & sharing

Feeds

Inline media

Workflow skill

See also

Resources

Errors