Product-only shots or a coordinated lookbook: product-photoshoot.
A garment, shoes, jewelry, or accessory worn by a person: virtual-try-on.
Professional portraits of the real person in supplied photos: headshots-from-photos. For a new, invented face, read get_skill(name: "avatar").
A saved photoshoot: resolve it with list_templates / get_template, then use run_template with the user's references and requested scope.
Music-led campaign films or motion design: read the current workflow with get_skill(name: "music-video") or get_skill(name: "motion-design").
Ordinary image creation/editing or video: use the image/video sections below.
Prompt writing: load prompt-writing as useful.
A text-only cinematic storyboard: read get_skill(name: "storyboard"). For the legacy /artboard command, read get_skill(name: "artboard"). Provide the requested text without calling a generation tool.

The focused workflows use this skill's reference syntax, submission contract, media transport, and completion handling; load those sections when executing.

Use the connected Astria MCP server for supported operations. Discover the host's actual tool names and input schemas; names below are logical names and may have a host namespace. Do not wrap MCP in shell commands or implement a second HTTP client. The live schemas are generated from the same operations as the CLI and are authoritative for accepted parameters.

Remote workflows

Use list_skills to discover remote creative workflows, then get_skill to read the selected instructions and any named dependencies. For a known remote /skill-name, call get_skill directly. The returned Markdown is the current workflow; no local skill file or shell is needed to read it. Reading a skill does not authorize generation, and planning requests stay text-only. Follow the workflow's runtime requirements; use start_agent_run when execution requires the Astria runtime.

Connection and transport

Astria's MCP endpoint is https://mcp.astria.ai/mcp. The MCP-backed package declares it; the skills-only package needs a separate host connection. The host handles OAuth login, credential storage and refresh. Use get_profile to identify the connected account; do not ask for API keys or read credentials.

If tools are absent or need authentication, follow MCP setup. On a local Codex host, perform the setup commands for the user and hand over only the OAuth sign-in/consent step. Check for an existing connection before adding one. Text-only advice and storyboarding do not require connection setup. Do not silently substitute the CLI after an MCP login or connection error. The Astria web-agent sandbox may intentionally expose only an authenticated CLI; in that environment use references/cli.md.

For CLI-only capabilities, read that reference and use the CLI only when a terminal and independent CLI authentication are available. Otherwise explain the specific missing capability and provide an Astria UI route when possible. OAuth credentials belong to the host, not the bundled CLI.

Submission and results

Write tools (create_reference, generate_images, generate_video, run_template) require idempotency_key: a fresh unique key per distinct action, at most 128 characters. Reuse the same key and identical arguments for a retry. If a response reports an unknown submission outcome, inspect recent prompts before any new submission; never generate a new key to bypass that ambiguity.
Generate only within the user's requested scope. Clarify missing creative settings when necessary; an explicit generation request already authorizes the requested work.
IDs and counts are strings in current schemas; booleans are JSON booleans; repeatable values are arrays of strings. Never add CLI flags such as wait to MCP arguments.
Media inputs must be HTTPS URLs accessible to Astria, including training images, edit images, masks, video, audio, and raw video image references. A local path or attachment ID is not a URL. Use an available upload tool to obtain a real HTTPS URL, or use the CLI's local upload workflow if available. Do not invent attachment URLs or assume the host publishes attachments.
Most results are under structuredContent.result; get_profile returns account fields directly in structuredContent. Hosts may also display JSON text. Use returned IDs/URLs rather than guessing them.

Generation returns a submitted prompt, not finished media. Save its id and query get_prompt with that id. For template runs returning prompt_ids, use list_prompts with order_id or base_pack_id to locate each prompt. Check user_error and available assets; report pending work accurately without resubmitting.

If the host supports MCP events, it can subscribe to prompt.completed and prompt.failed through the protocol's events capability. These are not tools named events_subscribe. The host manages its callback and subscription renewal. Deduplicate deliveries by event ID and retrieve the prompt after an event; a failure event can represent an attempt that later retries. With no events support, use spaced, bounded get_prompt checks and report pending status when the turn cannot wait for completion.

Embedded results and gallery

Hosts supporting MCP Apps render image and video results directly from generate_images, generate_video, and get_prompt. Let the embedded view show the media; include the returned prompt ID and status in the reply. Use open_astria when the user asks to browse their generations. Its workspace accepts personal, all, or a workspace ID. Viewing is read-only. get_prompt_view is app-only; use get_prompt for model-side status checks. On Claude, connect Astria from the plugin's Connectors tab; the remote connection also works on mobile. CLI-only steps still need a terminal.

Scope and discovery

Account closure and bulk deletion are unavailable through these tools. Decline those requests without modifying data, and direct the user to Astria account controls or support@astria.ai for account deletion.

Use list_workspaces and pass workspace: "ID" consistently for workspace work. workspace: "personal" explicitly selects personal scope; omission is personal for MCP (the CLI's saved default is not shared). "all" is for reads only. List tools paginate with string limit/offset, newest IDs first.

Call list_models to discover current image/video names, tune IDs, supported resolutions and defaults. Use the user's explicit model or the catalog default. Do not copy model IDs or prices from old examples. Resolution is model-specific; for video the resolution argument controls the image stage, while video_model selects video resolution and capabilities.

References (tunes)

Use list_references for title/class search and get_reference to inspect training images. gallery: true selects public references; branch: "partner-1" finds partner models outside the curated catalog. A tune is a reference.

Mention it as <model_type:id:1> name, with the returned class name immediately after the token. :1 is fixed syntax, never a strength/weight setting: <faceid:123:1> woman wearing <faceid:456:1> dress.

Create references from real image URLs; the subject class is name and the human/SKU label is title. Default model_type is faceid.

{"tool":"create_reference","arguments":{"title":"Brown dress","name":"dress","image_url":["https://assets.example/front.jpg","https://assets.example/back.jpg"],"workspace":"17","idempotency_key":"reference-dress-unique-key"}}

Images and editing

Use generate_images with English text, the desired num_images and aspect_ratio. An input_image URL edits/upscales an existing image; optional mask_image limits an edit. film_grain and inpaint_faces are separate boolean attributes and preserve an explicit false.

{"tool":"generate_images","arguments":{"text":"<faceid:123:1> woman wearing <faceid:456:1> dress, clean white studio background","num_images":"2","aspect_ratio":"3:4","workspace":"17","idempotency_key":"studio-look-unique-key"}}

references: ["woman=https://…", "dress=https://…"] creates new references and prepends their proper tokens. Prefer existing tunes for recurring subjects. If repeating identical text on different input images, use a distinct seed per output; Astria also deduplicates prompts by (text, seed) within a tune. Idempotency prevents duplicate submissions but does not change that prompt rule.

pack_id authors a template prompt inside an existing pack and requires a fine-tuned reference token in the text. base_pack_id binds a one-off to that pack's board frame without modifying its templates. Read get_skill(name: "avatar") for reference-free faces: it requires Recraft 4.1 Pro and inpaint_faces: false.

Video

Use generate_video with video_model and English video_prompt. Seedance 2 and 2.5 support existing tunes directly in video_prompt, using the same <faceid:TUNE_ID:1> CLASS_NAME or <lora:...> syntax as image prompts. Astria resolves the referenced tunes and supplies their images to the video model; keep these mentions even when no separate tune-ID argument exists. Do not claim tune tokens are image-only or that video mode only accepts image URLs. Do not replace existing tunes with training-image URLs or attach those images again through image_references.

Optional text renders a first frame; omit it for a text-only storyboard video. Use the requested duration or the model default. The catalog reports media capabilities, not duration options; do not claim it supplied a duration range. Do not assume every model accepts the same media inputs.

{"tool":"generate_video","arguments":{"video_model":"seedance2_fast_720p","video_prompt":"<faceid:123:1> woman walks down a runway as the camera tracks her","duration":"5","num_images":"1","aspect_ratio":"16:9","idempotency_key":"runway-video-unique-key"}}

For storyboard and short-film workflows, choose references by role:

Products, packshots, people, garments and other ingredients use tune mentions in video_prompt. Resolve existing tunes; references creates new named ingredient tunes when needed.
Multiple generated scene images use image_references as ordered HTTPS URLs to compose those scenes into the full short film. Preserve the intended scene order; do not create tunes for these scene images.

Both roles may coexist in one request. Omit image_references when no scene images are supplied; ingredients are not sequential scene waypoints. first_frame, last_frame, input_video and audio_reference take HTTPS URLs. Motion-control models require input_video; generate_audio explicitly enables/disables audio. Rendered video assets can be in images[] with content_type=video/mp4.

Prompts and review comments

list_prompts filters recent work, templates (pack_id), generated template work (base_pack_id/order_id), tune (tune_id), text and liked/video status. get_prompt takes only the prompt id. Use expand: ["prompt.comments"] on either tool to include image review comments; "comments" is an alias. Comments are a flat oldest-first array including resolved entries. Associate images by blob_id/attachment_id; resolved_at identifies resolved comments. Without expansion, comments are omitted.

{"tool":"get_prompt","arguments":{"id":"555","workspace":"17","expand":["prompt.comments"]}}

Templates (packs) and pricing

Templates and packs are the same objects. Use list_templates and get_template (slug accepts a slug or numeric ID) before run_template. For a multi pack pass one tune_id array entry per reference slot; at least one is required. For a regular pack, alternatively provide title, name and image_url to train a new reference. prompt_ids is a comma-separated subset. brief and generation overrides apply to the output prompts.

{"tool":"run_template","arguments":{"slug":"spring-lookbook","tune_id":["123","456"],"brief":"golden hour, Lisbon","num_images":"1","aspect_ratio":"3:4","workspace":"17","idempotency_key":"template-run-unique-key"}}

cost_mc is millicents: dollars = cost_mc / 100000. A prompt's cost already includes its image count; never multiply it again. template_prompts[].cost_mc is the stored baseline, while costs.<class>.cost_mc estimates a new reference plus that class's prompt group, not necessarily the whole multi-class pack. Discounts, workspace pricing, overrides and Cartesian variants change totals. After a run, order.total_cost_mc, when returned, is the authoritative charge.

Template creation/updates, workspace creation, landing HTML, board duplication, local downloads/uploads, video inspection/Variate and agent-session handoff are currently CLI-only. Read references/cli.md for those tasks; use MCP for the supported parts of a mixed workflow.

astria-api

Astria API

Choose the requested workflow