The Autodesk Fusion Compute MCP Server is in Public Beta — tools, supported operations, and behavior may change. Share feedback.
Your MCP client discovers tools automatically when it connects. Use this page as a quick reference, and use fusion_status — or your client's tool discovery — as the source of truth for what is available in your session right now.
The tools below are the top-level tools the server exposes. You use them to look up the API, manage a cloud Fusion session, and run work inside it.
| Tool | Description | When to use | What to provide | Example |
|---|---|---|---|---|
start_fusion |
Boots a headless Fusion instance in the cloud and returns its engine_session_id. |
First step before any modeling or CAM work. | Optional toolMode: Execute (default), CRUD, or CRUDE. Returns a queued status if capacity is full. |
{"name":"start_fusion","arguments":{"toolMode":"CRUD"}} |
fusion_status |
Reports session state and lists the worker tools currently available, plus your queue position when queued. | To confirm a session is ready, poll a queued start, or diagnose errors. | engine_session_id |
{"name":"fusion_status","arguments":{"engine_session_id":"..."}} |
invoke_fusion |
Proxy that lists, describes, or runs the session's worker tools. This is the tool that carries out modeling, CAM, document, and export operations. | To do the actual work in a running session. Call with no arguments to list available worker tools. | engine_session_id; tool and arguments when executing; optional async (see Long-running calls) |
{"name":"invoke_fusion","arguments":{"engine_session_id":"...","tool":"MCP.Read","arguments":{"queryType":"bodies"}}} |
fusion_call_result |
Returns the result of a call that is still running, waiting up to 40 seconds for it to finish. | When invoke_fusion returned status: "working" with a callId. |
engine_session_id and callId; optional waitSeconds (0–40) |
{"name":"fusion_call_result","arguments":{"engine_session_id":"...","callId":"..."}} |
stop_fusion |
Shuts down the instance and releases cloud capacity. Also cancels a queued start. | When you finish, or to recover a stuck session. | engine_session_id |
{"name":"stop_fusion","arguments":{"engine_session_id":"..."}} |
fusion_api_index |
Searches the Fusion TypeScript API index and the curated sample catalog. Sessionless — never starts Fusion or consumes capacity. | To confirm an API surface before scripting, or as a quick connectivity check. | query; optional source (api, samples, or all) |
{"name":"fusion_api_index","arguments":{"query":"extrude features","source":"api"}} |
fusion_get_sample |
Returns one TypeScript sample's full source and its prerequisites. Sessionless. | After fusion_api_index finds a sample, or when your client cannot read MCP resources. |
sampleId (the anchor from a search result's path) |
{"name":"fusion_get_sample","arguments":{"sampleId":"extrudefeatureapisample"}} |
fusion_sessions |
Lists the Fusion sessions you own and the available Fusion versions. Sessionless. | After reconnecting, to find a session that is still running, or as a quick connectivity check. | No parameters required. Optional filters (status, limit, and others). |
{"name":"fusion_sessions","arguments":{}} |
fusion_view_3d |
Exports the current design as STL or STEP and renders it inline in clients that support MCP Apps; other clients receive a browser viewer link. | To preview or share a design. | engine_session_id; optional format (stl or step) |
{"name":"fusion_view_3d","arguments":{"engine_session_id":"...","format":"stl"}} |
Beyond the top-level tools, a running session exposes a set of server-side worker tools that carry out the actual Fusion operations. These are not called directly — they are reached through invoke_fusion, grouped into MCP.* namespaces.
Under the default Execute tool mode, the worker surface is MCP.ExecuteTypeScript plus MCP.IO, MCP.Documents, and MCP.FusionTeam — with advanced work performed by running TypeScript against the Fusion API. Pass toolMode: "CRUD" to start_fusion instead to get the handcrafted native tool set, including the manufacturing and assembly namespaces; CRUDE combines the native tools with MCP.ExecuteTypeScript.
The exact set varies with the Fusion build behind your session and with server configuration. Call invoke_fusion with no arguments (list mode), or fusion_status, to see exactly what your session offers.
A worker tool call that takes longer than about 40 seconds does not fail and does not need to be repeated. Instead, invoke_fusion returns status: "working" with a callId, and the call keeps running in Fusion. The tool descriptions tell agents to collect the result with fusion_call_result.
fusion_call_result waits up to 40 seconds and returns as soon as the call finishes. The result is exactly what invoke_fusion would have returned, including an error if the call failed. If the call is still running, it returns status: "working" with the same callId again; call it again.invoke_fusion call again would run it a second time.invoke_fusion or fusion_view_3d call sent while one is running is rejected with the running call's callId. fusion_status, fusion_call_result, stop_fusion, invoke_fusion list mode (no arguments), and the sessionless tools keep working.callId at once. Pass "async": true to invoke_fusion to get the callId back immediately, for a call you expect to run long. This is not available for MCP.IO screenshots with delivery: "signedUrl" or for fusion_view_3d, which wait for their export to finish.stop_fusion also ends a running call, which is then reported as failed.callId. If your client lost the callId, for example because its own request timed out, fusion_status shows the running call as pendingCall.{"name":"invoke_fusion","arguments":{"engine_session_id":"...","tool":"MCP.ExecuteTypeScript","arguments":{"code":"..."},"async":true}}
This server exposes one MCP resource, which your client discovers automatically:
| Resource | Description |
|---|---|
mcp://fusion/samples/typescript-script-samples |
A catalog of Fusion TypeScript samples, grouped by category. Each code block is a starting point for MCP.ExecuteTypeScript — adapt it to your task and check its stated prerequisites, since some samples need an existing document, CAM setup, or cloud file. |
The same samples are searchable through fusion_api_index with source="samples", and fusion_get_sample returns one sample's source directly if your client cannot read MCP resources.
MCP.ExecuteTypeScriptInline scripts run as bare scripts, not modules, so a top-level import or export is rejected. You do not need one: adsk is already in scope as a value. Where you need a type by name, declare a local alias from a type query.
type Design = import("@adsk/fusion").adsk.fusion.Design;
const app = adsk.core.Application.get();
const design = app.activeProduct as Design;
adsk.result = JSON.stringify({ name: design.rootComponent.name });
Spell adsk.log and adsk.result literally. Output capture is a textual rewrite, so an alias or adsk["result"] still runs but its output is not returned to you.