Skip to content
Docs menu

Studio for agents

The studio_* MCP tools and the REST API, the cheap workflow that keeps an agent's token use small, and what each answer returns.

View .md

Studio is a product of the workspace MCP server. Connect it alone, so its tools are the only ones in your agent's context:

claude mcp add --transport http venom-studio https://cloud.usevenom.dev/workspace/mcp/studio

The scopes are studio:read and studio:write. A personal key or an OAuth grant acts as its member by their role; a shared key bound to a service agent acts as that agent; a bare shared key only reads what the organisation shares. In the agents portal, turn on the Venom Studio server in a quick chat.

How an artifact gets made

For a report that needs research, an agent can hand the whole process to the server: studio_research_start runs goals, parallel researchers, an analysis per section, the build and a review, and studio_research_get shows its plan card (Research runs). A quick chat does this by itself for such requests. The skill below is the process for everything else, and for clients that build the artifact themselves.

Every agent follows one process, the studio-artifact skill: a short brief, research where the topic needs facts the chat does not have, the outline written in the chat before anything is built, the build from the catalog, a review against the plan, and an answer with the url, the assumptions, the sources and what is still unverified. It checks the request's own premise too: asked about "the 1987 Plaza Accords", it says the accord was signed in 1985 and builds on that.

  • In a quick chat with Venom Studio in the picker, the skill is offered even when the member picked none, and Studio's guidance tells the model to load it before it makes anything. A model that skips it anyway is walked through it: its first studio_artifact_create in the chat, made before the skill was read, does not run. Its answer is the skill's steps, headed "Not run yet: follow this first", and the chat shows it as the skill being loaded. The same create runs when the model calls it again, after its brief and outline. This happens at most once per chat, never after the skill was loaded or typed as /studio-artifact, and not when an admin disabled the skill.
  • Any other MCP client reads the same text with studio_catalog_get and group: "workflow".
  • In Claude Code, the Venom plugin ships it as /venom:studio-artifact.
  • It is in every organisation's Skills registry, where an admin can scope, edit or disable it like any other skill.

For research it uses what the chat has: workspace_web_search (Venom Search, which a quick chat offers on the Studio server, off until an owner or admin turns it on in Settings, Search), any other search tool the member added, the organisation's notes, issues and people, and what the member pasted or linked. With no search at all it still works: figures it could not check are marked approximate and listed under "To verify".

Stubs are refused

An agent's write that is plainly a stub is not saved. The answer says what is thin and how to fix it. This applies to studio_artifact_create, studio_content_edit and studio_html_edit from a key or an OAuth grant (MCP and REST); a person in the console is never checked, and nothing already saved is touched. A stub is:

Kind Stub when
HTML page under 600 characters of visible text with nothing else on it (no script, SVG, canvas, form, table or image, fewer than 25 elements); or most paragraphs end in "..."
deck fewer than 3 slides, or fewer than 2 slides with anything beyond a title
page fewer than 3 blocks, only headings, or under 200 characters beyond the headings with no stats, chart, diagram, table or image
any placeholder text: "lorem ipsum", "content goes here", "[insert ...]", "placeholder text", "TODO:"

Pass draft: true to save one anyway: the first part of a long build, or something meant to be small, like a title card or a single tile. The answer then ends with Draft: thin, not finished: and the list, and structuredContent.draft carries the findings. Adding to something that is already thin needs no flag, so building in parts works; only a write that makes a finished artifact thin, or replaces one whole with a stub, is refused. studio_artifact_validate reports the same findings (thin, findings) without saving.

A slide is a fixed frame, so text that does not fit is cut at its edge. When an agent creates or edits a deck, the answer ends with Too much on ... and names each slide that holds more than fits, by number and id, with how much fits there (about 400 characters on a bullets slide, 520 across two columns, 7 table rows counting the caption and source). The deck is saved anyway; the agent shortens the points, splits the slide, or moves detail to the notes. An agent's create of a deck or page without content is refused with what is missing and an example.

The cheap workflow

  1. studio_catalog_get: the block types and layouts, one line each. studio_component_get gives the schema and an example of the ones you will use; you never carry the whole library.
  2. studio_artifact_create with the whole outline: a deck's slides or a page's blocks. The answer lists the ids. When the artifact is too long for one answer, create it with the first part and add the rest with studio_content_edit (or studio_html_edit for an HTML page).
  3. studio_outline_get, then studio_content_edit by id, one thing at a time. Read the outline, not the artifact.
  4. studio_artifact_validate to dry-run a spec you are unsure of (it also says whether it is a stub), and dry_run on an edit.
  5. studio_preview_get for the view, present and print links and a plain-text summary of each slide, to check the result without images.
  6. Give the user the url of every answer. Share outside the organisation with studio_share_create.

The server's instructions, which every MCP client shows its model and a quick chat puts under Studio in its system prompt, add when to pick each kind (a deck to present, a page to read, an HTML page for mockups, prototypes, landing pages and interactive work) and what good work looks like: an outline with a point per slide, takeaway headlines, the real numbers from the conversation, data shown as charts, stats and tables, and for HTML a self-contained, responsive page with a deliberate type scale, one accent colour and light and dark modes. They tell the model to use the org's theme for HTML by default (<meta name="venom-theme" content="org">, themed HTML pages); studio_catalog_get with group: "html" lists the kit's variables and classes.

An HTML page is shown sandboxed with no network. When studio_artifact_create, studio_html_edit or studio_artifact_validate sees a CDN script, a <link> stylesheet, a web font or a remote image, the answer names each one and says to inline it; the page is still saved.

Tools

Tool Does
studio_catalog_get, studio_component_get, studio_layout_get the component library on demand
studio_artifact_list, studio_artifact_search find artifacts (yours, and what is shared with you)
studio_outline_get, studio_artifact_get read: ids and titles only, or fields and ids to read part
studio_artifact_create, studio_artifact_update, studio_artifact_duplicate, studio_artifact_delete, studio_artifact_restore the artifact (title, theme, its own tweaks over the theme: Tweaks, sharing)
studio_content_edit, studio_html_edit anchored edits, in batches
studio_artifact_validate, studio_preview_get, studio_artifact_export check, look, and take it out as json, markdown, outline, html or pptx (a deck as a PowerPoint file, base64)
studio_version_list, studio_version_restore history
studio_share_list, studio_share_create, studio_share_revoke public links
studio_theme_list, studio_theme_set the organisation's themes
studio_theme_import a proposed theme from a design system's files, saved only with save: true (Import a design system)

Every write answers with the artifact's url, its new version, and the ids it touched, in the text and in structuredContent. Creates take an idempotency_key, so a retry after a timeout returns the first result. Lists page with cursor and limit. expect_version on a write fails it when someone else wrote first.

Refusals name the path and the allowed values, all at once, and never write anything:

The edit is not valid; nothing was changed.
- ops[0].content.titel: is not a slot of the bullets layout, did you mean "title"? Slots: title, bullets.

A few guesses are taken as meant instead of refused: artifact_id for id; components, names, type or id for studio_component_get's types; layout or layouts for studio_layout_get's id; and a create's slides or blocks sent next to kind instead of inside content. The lookups (studio_catalog_get, studio_component_get, studio_layout_get) never refuse over an argument: one they do not take is ignored, a list where they read one value gives its first, and more than six types gives the first six. The answer starts with a line that says what was ignored or cut.

REST

The same functions are under /api/v1/studio, with an API key as Authorization: Bearer:

Method and path Tool
GET /studio/catalog, /studio/components?types=bar,table, /studio/layouts the library
GET /studio/artifacts, POST /studio/artifacts list, create
GET, PATCH /studio/artifacts/{id} read, update (If-Match is expect_version)
GET /studio/artifacts/{id}/outline outline
POST /studio/artifacts/{id}/edit, /html edit slides or blocks, edit HTML
POST /studio/artifacts/{id}/delete, /restore, /duplicate trash, restore, copy
GET /studio/artifacts/{id}/export, /preview, /versions export, preview, history
POST /studio/artifacts/{id}/versions/{version}/restore restore a version
GET, POST /studio/artifacts/{id}/shares, POST .../shares/revoke public links
GET /studio/themes, PUT /studio/themes/{id} themes
POST /studio/artifacts/validate, GET /studio/artifacts/search dry run, search

A stale If-Match is a 412 version_conflict; the other refusals are 4xx with the same words as the MCP tools.