@marlinjai/studio-mcp 0.4.0 → 0.6.0

package/README.md

@@ -2,7 +2,7 @@
 
 A thin [Model Context Protocol (MCP)](https://modelcontextprotocol.io) server over the **hosted Lumitra Studio HTTP API** (`studio.lumitra.co`). It gives any MCP client (Claude Code, Cursor, Claude Desktop, your own agent) typed tools for brand-consistent image generation, character minting, workflow authoring and runs, published recipes, and the still-swap video recipe, without cloning the Studio repo, without Postgres, without provider keys. The server is a pure HTTP client: every tool maps onto an existing `/api/**` route.
 
-Every generation spends real money on the Studio's provider accounts (roughly $0.04 per image, $0.30 per character mint, $0.70 per video take). Treat the API key like a production credential.
+Every generation spends real money (on prepaid credits without a subscription roughly $0.06 per image, $0.84 per character mint and $1.05 per video take; the exact amount is the `costUsd` every job reports). Treat the API key like a production credential. A key belongs to one company: everything it reads, creates and spends is that company's.
 
 ## Install
 
@@ -23,6 +23,10 @@ claude mcp add lumitra-studio \
 
 If `LUMITRA_STUDIO_API_KEY` is already exported in your shell, drop the `--env` flag: the server reads it from its own environment.
 
+**Changed in 0.6.0.** `costUsd` and every spend figure are now what the job cost YOUR company: for a company on prepaid credits the amount that left its balance, where 0.5.0 and older described a provider figure. See [What changed in 0.6.0](#what-changed-in-060).
+
+**New in 0.5.0.** `studio_generate_image` no longer requires a `brandSlug` (leave it out for an image without a brand), takes a `projectId`, and returns the Studio's `candidates` plan and `notice`. Nothing breaks for a caller that sends a real brand slug. See [What changed in 0.5.0](#what-changed-in-050).
+
 **Breaking in 0.4.0.** The legacy variable `LUMITRA_STUDIO_SERVICE_TOKEN` is no longer read. Provide the same key as `LUMITRA_STUDIO_API_KEY`; a server started with only the legacy name stops with an error that says so.
 
 ### Cursor (`.cursor/mcp.json`) and Claude Desktop
@@ -79,24 +83,25 @@ The token is held in memory only. It is never logged, never serialized into a re
 
 | Tool | What it does |
 |---|---|
-| `studio_list_projects` | List projects (id, slug, name). A `projectId` is needed by the workflow-based tools. |
-| `studio_generate_image` | Submit a brand-styled image generation (optionally pinned to a character), wait for it, return the signed URL and cost. `wait=false` returns the `jobId`. |
-| `studio_generate_3d` | Generate one 3D model (a textured GLB mesh) from a prompt or one image through 3D AI Studio, wait for it, return the signed GLB URL and cost (typically $0.25 to $0.70 per model). Refuses a model and image mismatch before submitting. `wait=false` returns the `jobId`. |
+| `studio_list_projects` | List projects (id, slug, name). A `projectId` is needed by the workflow-based tools and optional on `studio_generate_image` and `studio_generate_3d`. |
+| `studio_generate_image` | Submit an image generation, wait for it, return the signed URL and cost. Only `prompt` is required: `brandSlug` (a brand of the key's company) applies that brand's style and references, and leaving it out generates without a brand; `characterSlug` pins a character; `projectId` chooses the project the image is filed in. A `brandSlug` or `characterSlug` the company cannot use is refused with `404` `input_not_found` before anything is spent. The result carries `candidates: { requested, rendered, mayRender? }` and, when the model renders fewer images than `candidateCount` asked for, a `notice`. `wait=false` returns the `jobId`. |
+| `studio_generate_3d` | Generate one 3D model (a textured GLB mesh) from a prompt or one image through 3D AI Studio, wait for it, return the signed GLB URL and cost (the price follows the model and the quality flags). Refuses a model and image mismatch before submitting. `wait=false` returns the `jobId`. |
 | `studio_rig_3d` | Rig a humanoid Model3D and apply up to 10 Meshy animation presets (by name or id) through a one-node workflow; returns the rigged GLB, each clip's name and url, and the cost ($0.20 per rig plus $0.12 per preset). Refuses an unknown preset, with close matches, before submitting. |
 | `studio_find_animations` | Search Meshy's 678-preset animation library by words. Local, free, no request. |
 | `studio_get_job` | Read a job's status, cost, error and fresh result URL; `wait=true` blocks until terminal. |
-| `studio_list_characters` | List minted characters. |
-| `studio_get_character` | One character with its frozen reference images (category + signed URL). |
+| `studio_list_characters` | List the characters the key can use: its company's own, plus the ones another company shared with everyone (`owned: false`, usable in generations, read-only). A slug is unique within one company. |
+| `studio_get_character` | One character with its frozen reference images (category + URL). For a character the company owns the URL is a direct image URL; for one another company shared (`owned: false`) it is a Studio address that works unchanged as an image input to any Studio tool, is not a public image URL, and stops working when the owner stops sharing. |
 | `studio_create_character` | Create an empty character record (for your own references). Requires `likenessRightsConfirmed: true`. |
 | `studio_mint_character` | Run the curated character-sheet workflow inline, wait, create the character, upload the 7 views as references. Requires `likenessRightsConfirmed: true` and refuses before the run starts without it. Returns `slug`, `referenceCount`, `costUsd`, `runId`. |
 | `studio_create_brand` | Create a brand (style prefix + context). |
-| `studio_upload_brand_reference` | Upload a reference image to a brand library from a local path or URL. |
+| `studio_upload_brand_reference` | Upload a reference image to a brand library from a local path or URL. The `brandSlug` must name a brand of the key's company (anything else is a `404`). |
 | `studio_run_workflow` | Run an inline workflow DAG (or a saved one by `savedId`), with params, node subsets, repeat count and `visibility`; waits by default. |
-| `studio_get_run` | Read a run: overall status, per-node status, URLs, cost, errors. |
+| `studio_get_run` | Read a run: overall status, per-node status, URLs, cost, errors. A run parked on a person at a judge node comes back with a `waiting` block naming the node and its candidates. |
+| `studio_decide_judge` | Submit a PERSON'S approve or reject on a parked judge node; approving resumes the run. Never the agent's own judgement. |
 | `studio_get_run_output` | Fresh signed URL for one node's output. |
 | `studio_replace_character_in_video` | The proven still-swap recipe: image-edit the character into a scene frame, then image-to-video. Accepts `characterSlug` (full-body reference auto-resolved) or `characterReferenceUrl`, optional `endFrameUrl` (switches to Kling start+end keyframes). Returns `stillUrl`, `videoUrl`, `costUsd`, `runId`. |
-| `studio_get_spend` | Spend ledger: today / 7-day / 30-day totals plus recent paid actions. |
-| `studio_get_usage` | This key's tenant against its plan: month-to-date spend and job count, the monthly USD cap and requests-per-minute limit (`null` = unlimited), and a per-model breakdown. Call before a large batch to check remaining budget, or after a `402`/`429` to see what tripped. `usedUsd` includes in-flight (not yet settled) job estimates, so it can go DOWN as well as up: a job that later fails releases its estimate. |
+| `studio_get_spend` | Spend ledger: today / 7-day / 30-day totals plus recent paid actions, in what the company paid. |
+| `studio_get_usage` | What limits the key's company and what it has used: month-to-date spend and job count in what the company paid, the monthly USD cap (`capUsd`) and requests-per-minute limit (`null` = none), a per-model breakdown, and `credits`. For a company on prepaid credits (every new company) `credits` is `{ balanceUsd, dailyLimitUsd, usedTodayUsd, dailyResetsAt, operatorMonthlyLimit }`: the balance, the daily limit, what left the balance today, when the day resets and whether the operator set a monthly limit on top; for a company held to a monthly cap it is `null`. Call before a large batch to check remaining room, or after a `402`/`429` to see what tripped. `usedUsd` includes in-flight (not yet settled) job estimates, so it can go DOWN as well as up: a job that later fails releases its estimate. |
 | `studio_workflow_describe` | The node vocabulary: every task type with ports, prompt policy and runnability, the input types, the op types, the patch limits. Settings and models only with `includeSettings` / `includeModels`; filter with `tasks`. |
 | `studio_workflow_get` | A saved workflow as a summary (nodes, wire counts, edge and group counts, `revision`, `readOnly`, `projection`), or the whole response with `full: true`. |
 | `studio_workflow_create` | Build a new saved workflow from ops applied to an empty document. Returns `{ ok, savedId, revision, summary }`. |
@@ -112,6 +117,13 @@ All waiting tools accept `waitSeconds` (default 600) and back off between polls
 ### Tool signatures (the contract shared with the `lumitra-studio` plugin)
 
 ```
+studio_generate_image
+  in : { prompt, brandSlug?, brandMode?, characterSlug?, projectId?, model?, aspectRatio?, imageSize?,
+         seed?, candidateCount?: 1..4, wait?: true, waitSeconds? }
+  out: { jobId, status, model, costUsd, errorMessage, url, assetId, urls[], kind,
+         candidates?: { requested, rendered, mayRender? }, notice? }
+       wait=false: { jobId, kind, status: "queued", candidates?, notice? }
+
 studio_mint_character
   in : { projectId, name, slug, description, likenessRightsConfirmed: true, descriptorPrompt?,
          negativePrompt?, defaultModel?, seedModel?, viewModel?, visibility?: "listed"|"unlisted",
@@ -127,6 +139,24 @@ studio_replace_character_in_video
 
 Reference categories written by the mint: `face`, `three-quarter`, `profile`, `full-body`, `close-up`, `lighting-warm`, `lighting-cool` (the Studio's canonical names, the same ones its in-app mint flow and try-on ranker use).
 
+### What changed in 0.6.0
+
+No argument and no result field was renamed or removed. What changed is the meaning of the amounts the Studio returns to a company on prepaid credits, since 2026-10-11, and the text of the tools that describe them:
+
+- **`costUsd` is what the job cost your company.** For a company on prepaid credits it is the amount that left its balance for the job (the same figure the estimate before a run shows), on `studio_generate_image`, `studio_generate_3d`, `studio_rig_3d`, `studio_get_job`, `studio_get_run`, `studio_run_workflow`, `studio_run_published`, `studio_mint_character` and `studio_replace_character_in_video`. A program that added a factor of its own to `costUsd` to predict the balance must stop: the balance falls by `costUsd`. For a company held to a monthly cap arranged with the Lumitra operator nothing changes.
+- **`studio_get_spend` and `studio_get_usage` report what was paid.** For a company on prepaid credits the totals of `studio_get_spend` equal what left the balance for the jobs of each window, and `usedUsd`, `realizedUsd`, `inFlightEstimatedUsd` and `byModel` of `studio_get_usage` are in the same dollars as `credits.balanceUsd`. `capUsd` is always `null` on credits; the new `credits.operatorMonthlyLimit` says whether the Lumitra operator set a monthly limit on top, and a `monthly_cap_reached` refusal of such a company carries `resetsAt` and no amounts.
+- **The typical prices in the descriptions are prices.** The figures in the tool descriptions and the skill are Studio's prices for a company on prepaid credits without a subscription (a character mint is about 0.84 USD, where the text used to say about 0.30 USD).
+
+### What changed in 0.5.0
+
+Nothing breaks for an existing caller that sends a real brand slug: every argument that existed keeps its name and meaning, and every result keeps its fields.
+
+- **`studio_generate_image` no longer requires `brandSlug`.** Leave it out and the image is generated without a brand. Up to 0.4.0 the field was required, so an agent of a company with no brand sent a made-up slug. Since 2026-10-11 the Studio refuses a slug that names no brand of the key's company (`404`, `code: "input_not_found"`) instead of generating without a style into a project of that name, which left such an agent with no way to make an image at all.
+- **`studio_generate_image` takes `projectId`.** The image is filed in that project (one of the company's, from `studio_list_projects`). Without it the image lands in the project named after the brand, or in `default` when there is no brand.
+- **The candidate plan is passed through.** The result carries `candidates: { requested, rendered, mayRender? }` as the Studio reported it when it accepted the job, and a `notice` sentence when fewer images will be made than `candidateCount` asked for: most KIE image models render one image whatever the count. Only rendered images are charged.
+- **A refusal reaches the agent as the Studio worded it.** The Studio's `error` sentence, which names the next step, is handed on unchanged, followed by a `hint` the sentence does not already contain and by the machine `code`, for example `POST /api/generate -> 404: Brand 'acme' not found. Create the brand first (POST /api/brands, or the studio_create_brand tool), or leave brandSlug out to generate without a brand. (code: input_not_found)`. Up to 0.4.0 the `code` and the `hint` were dropped. A client that matched a tool error by its exact full text sees the added ` (code: ...)` suffix; the sentence before it is unchanged.
+- **Descriptions match the Studio of 2026-10-11.** `studio_get_usage` describes the `credits` field (balance, daily limit, used today, reset time); `studio_list_characters`, `studio_get_character` and the `characterSlug` arguments describe shared characters (readable and castable, not editable, reference images as Studio addresses); `studio_upload_brand_reference` and `studio_replace_character_in_video` say that a slug the company cannot use is a `404` before anything is spent.
+
 ### The likeness confirmation (every tool that creates a character)
 
 `studio_create_character` and `studio_mint_character` require `likenessRightsConfirmed: true`, and `studio_run_published` requires it for a recipe that creates a character. It records, on the character, that the human the agent acts for confirmed:
@@ -194,6 +224,11 @@ Generate a 16:9 image for brand "busbasis" with Bruno at a workbench in a bright
 model kie/nano-banana-2, and tell me the cost.
 ```
 
+```text
+Make one image of a lighthouse at dusk, no brand, and file it in the project "moodboards".
+How much credit do we have left afterwards?
+```
+
 ```text
 Here is a frame from our header video: https://.../frame-012.png
 Put bruno-hartmann into that shot and give me the clip. Use the full-body reference.