AI assistants (MCP)

Prism runs an MCP server, so an AI assistant can search your projects, read components, nets and BOMs at a revision, compare revisions, render views, leave review comments and add parts to the component catalog. It acts as you, with your role and folder visibility, and it can’t sign anything off.

Turn on the MCP server

The MCP server is off by default. An administrator turns it on:

1. Set PRISM_MCP_ENABLED=true.
2. Set PUBLIC_BASE_URL to the address people use to reach Prism, for example https://prism.example.com. Sign-in tokens are bound to <PUBLIC_BASE_URL>/mcp, so it must match what clients connect to.
3. Restart the backend.

Both settings are described in Configuration. While the server is off, /mcp and its sign-in metadata answer 404.

Caution

Whatever an assistant reads is sent to that assistant’s model provider. Before you turn MCP on, decide which projects may leave Prism this way, and limit MCP to them with Project scope.

Connect Claude Code

1. Add the server:

   claude mcp add --transport http prism https://prism.example.com/mcp

2. Start Claude Code and ask it something about a project, for example “List the Prism projects”. The first call opens Prism in your browser.

3. Sign in if asked, check the access the client asks for, and approve it.

The client is now connected and appears under Settings → Connected MCP clients.

Here Claude Code compares two releases of an open hardware board through Prism, and summarises what changed:

Connect other clients

The server address is https://<your Prism>/mcp. It speaks MCP over Streamable HTTP and signs clients in with MCP OAuth, so a client that supports OAuth finds the sign-in page on its own.

• Claude Desktop and claude.ai: add a custom connector with the server address.

• A client without OAuth: send a personal access token as the header Authorization: Bearer prism_pat_.... In Claude Code:

  claude mcp add --transport http prism https://prism.example.com/mcp --header "Authorization: Bearer prism_pat_..."

A browser session cookie isn’t accepted at /mcp.

Approve access

A client signed in with OAuth starts with read access (api:read) only. The first time the assistant tries a write, Prism asks your client for just the scope that write needs, and you approve it in the browser. The scopes a client can be granted are api:read, comments:write, design:write and catalog:write, plus ops:read and ops:act for administrators when ops diagnostics are on.

A client using a personal access token isn’t asked. A write its token can’t make fails with the code insufficient-scope, naming the scope it lacks.

Read tools

These tools change nothing in Prism.

Tool	Returns
render_view	a PNG of a schematic sheet, a PCB layer set (front, back, copper or all) or a 3D view (top or bottom) at a revision, with chosen components or nets highlighted, optionally cropped to a region in mm. 3D views have no highlights. A render still running returns its status; ask again with the same arguments.
search_projects	projects matching a name, description or repository
get_project	one project’s record and latest sync
list_revisions	a project’s commits, newest first
find_in_design	components and nets at a revision matching a reference, value, field or name
get_component	one component at a revision: value, footprint, sheet, assembly state, fields, and pins with their nets
get_net	one net at a revision: pins, connected components, class and aliases
get_bom	the engineering BOM at a revision, grouped as the UI groups it, all lines or assembly only
get_comparison_review	the review groups of a finished comparison for the schematic or the PCB
get_job	a job’s status, stage and message
list_comments	a project’s comment threads, optionally only open or resolved ones
search_catalog	catalog components matching a part number, manufacturer or description
get_catalog_component	one catalog component with its assets, workflow stage and release state
list_catalog_fields	the catalog’s metadata fields, including custom ones
search_catalog_assets	symbols, footprints and 3D models the catalog already stores
check_catalog_asset	whether KiCad can read a draft symbol or footprint, its KLC findings and a PNG of each unit, without storing anything. It needs catalog:write.
get_release_status	a project’s Release Studio candidates and their builds
compare_revisions	starts a comparison of two commits, branches or tags and returns its job id; it changes no design
get_changes	the review rows of a finished comparison, filtered by kind (added, changed, removed), impact or reference, one page at a time

To review a change, an assistant calls compare_revisions, polls get_job until the job is completed, then reads get_changes. A client that supports the MCP tasks extension gets a task for compare_revisions instead, and follows it with tasks/get.

Write tools

Tool	Does	Scope
sync_project	queues a sync of the project with its remote; returns a job id	design:write
start_release_build	queues a Release Studio candidate build for one commit; returns a job id. It never approves or publishes.	design:write
request_release_approval	changes nothing: returns the project’s Release Studio page, where a person approves	design:write
create_catalog_draft	creates an unreviewed catalog component in the open workflow stage	catalog:write
update_import_proposal	saves draft edits to a catalog import session’s proposals; it doesn’t accept, reject or release anything	catalog:write
update_catalog_component	fills in or corrects a component’s metadata	catalog:write
create_asset_upload	returns a one-time URL to PUT a file too large to send inline, such as a STEP model	catalog:write
attach_catalog_asset	attaches a symbol, footprint, 3D model or SPICE model, inline or from an upload	catalog:write
reuse_catalog_asset	gives a component an asset the catalog already stores, shared rather than copied	catalog:write
validate_catalog_component	runs KLC validation and returns the findings	catalog:write
update_catalog_representation	points a representation (a symbol and footprint pair) at other assets, renames it or makes it the default	catalog:write
remove_catalog_representation	removes a representation and detaches its assets when nothing else uses them	catalog:write
move_catalog_component	moves a component to In progress or QA Review, never further	catalog:write
request_catalog_approval	changes nothing: returns the component’s page, where a QA reviewer approves it	catalog:write
add_comment	comments on a component, a net, a schematic sheet or one comparison change, with an optional class and severity. Prism places the comment from the design at that revision, not from coordinates.	comments:write
reply_to_comment	replies in an existing thread	comments:write
resolve_comment	changes nothing: returns the Prism page where you resolve the thread yourself	comments:write

resolve_comment, request_release_approval and request_catalog_approval hand the step back to you as a link. Your client opens it, and you act in Prism.

Add a part to the catalog

With the catalog tools an assistant can take a new part from an empty record to QA Review: create it, fill its fields, check and attach a symbol and footprint (or reuse a stored footprint), attach a STEP model, validate, and move it to QA Review. Prism doesn’t make the files. The assistant brings them, and Prism checks and stores them. Prism embeds an attached 3D model in the footprint it serves.

A corrected file attached later adds a representation. The assistant then replaces the old one with update_catalog_representation and remove_catalog_representation. Approving the part stays with a QA reviewer.

Every write tool takes an optional idempotencyKey. A client that retries a write with the same key gets the first result back, not a second comment.

What results contain

• Each design result names the full commit it read. Pass that commit as revision to keep later calls on the same design.
• Results cite a Prism link that opens the viewer on exactly that component or net (see Links to a view) and a prism:// resource URI.
• The server also offers components, nets and BOMs as resources, at prism://project/{project_id}/rev/{commit}/component/{key}, .../net/{key} and .../bom.
• Every result is marked as untrusted data. Comment text, field values and file contents are returned as data, and no tool acts on instructions found inside them.

Comments and replies an assistant writes through OAuth carry an Agent badge in Prism. Hover it to see the client’s name.

Control what assistants reach

Administrators configure MCP under Settings → MCP:

• Project scope: All projects (each person still sees only what their role and folder visibility allow) or Listed projects. A project outside the list looks to an assistant exactly like one that doesn’t exist.
• Tools: switch any tool off. A disabled tool disappears from the assistant’s tool list.
• Audit: every MCP tool call and resource read is recorded with the time, user, client, token, tool, project, revision and outcome, and so is every write made through the HTTP API with a token. Filter by client, token, project, tool, or reads and writes. How long rows are kept is set in Configuration.

Each person sees their connected clients, with the scopes granted and last use, under Settings → Connected MCP clients, and clicks Revoke to disconnect one. That page shows only your own clients. An administrator lists everyone’s with GET /api/v1/auth/mcp-clients/all and can revoke any of them with DELETE /api/v1/auth/mcp-clients/{grant_id}.