TokenLab Skills

Agent Skills for TokenLab. Each skill is a recipe for one outcome: a product photo, an animated still, a voiceover, a 3D asset, a typed decision. The agent loads a skill from its description, follows the workflow, and calls the TokenLab API over HTTP or MCP. It reads the live model catalog at run time, so the skill stays correct when models change.

How the skills fit together

• Foundation skills do the shared work:
  • tokenlab-run: the execution contract. Read the model contract, estimate the cost, provide inputs, call, wait for the task, and read the result.
  • tokenlab-model-picker: map an outcome to an operation and to live model IDs.
  • tokenlab-prompting: prompt craft per model family.
• Outcome skills (image, video, audio, 3D, decisions) name default models, a workflow, technique, and a quality bar. They call tokenlab-run for the mechanics.
• Integration skills put TokenLab into application code: SDK setup, native endpoints, cost routing, and migration from other OpenAI-compatible providers.

Install tokenlab-run together with any outcome skill. The plugin installs all of them.

Catalog

catalog.json owns the skill list, groups, order, operations, and preferred models. skills.sh.json and this section are generated from it with python3 scripts/sync_catalog.py.

Model defaults refresh daily

Skills do not hard-code model IDs in their text. Each skill's ## Models section holds a block that scripts/sync_models.py generates from the live catalog (GET https://api.tokenlab.sh/v1/models, no key needed):

1. The skill's preferredModels from catalog.json that are live and support the operation come first.
2. If none is live, the top three models that recommended_for=<scene> ranks by 24-hour success rate take their place, plus one low-cost option. When the API does not offer a scene yet (for example decision), the category=<scene> list is used instead.

The script also writes generated/availability.json: for each skill, status (available, or unavailable when no live model supports any of its operations), the current models, and checkedAt. Capability-specific operations (for example background removal, vectorizing, avatars, upscaling, motion control) set rankedFallback to false in catalog.json: when their preferred models are gone, the skill shows the operation as unavailable instead of a general model. Deprecated and preview models are never listed.

The Sync model defaults workflow runs it daily:

• If a catalog read fails or the catalog lists fewer than 200 models, it stops and writes nothing.
• A small change (at most 3 skills, and at most 2 that become unavailable) gets a patch version, passes the tests and skill validation, and is pushed to main. Plugin and Skills CLI users get it on their next update.
• A larger change opens a pull request for review instead.

Foundation - Run any TokenLab request correctly, pick the model, and write the prompt.

• tokenlab-run - The execution contract: contract, estimate, upload, call, poll, read results.
• tokenlab-model-picker - Pick the model for an outcome from live catalog signals.
• tokenlab-prompting - Prompt craft per model family for image, video, music, and speech.

Image - Create and edit images: product shots, edits, upscales, text, characters, photoreal stills, and logos.

• tokenlab-product-photography - Packshot to studio, lifestyle, and hero shots.
• tokenlab-edit-image - Add, remove, replace, recolor, and cut out backgrounds.
• tokenlab-restore-and-upscale - Upscale, sharpen, and restore old or low-resolution photos.
• tokenlab-text-in-image - Posters, packaging, and UI mockups with exact, legible text.
• tokenlab-character-consistency - Keep one character or product the same across images and clips.
• tokenlab-photoreal-stills - Realistic photos of people, places, food, and scenes.
• tokenlab-logos-and-vectors - Logos, icons, and flat marks, with SVG vectorizing.

Video - Animate stills, make ads and cinematic shots, control frames and motion, extend and edit clips.

• tokenlab-animate-image - Turn a still image into a short video clip.
• tokenlab-product-ad-video - Product image or brief to a short ad clip.
• tokenlab-keyframe-transition - Video that moves from a first frame to a last frame.
• tokenlab-extend-video - Continue a clip past its last frame.
• tokenlab-edit-video - Restyle or change an existing clip with a prompt.
• tokenlab-motion-control - Move a character image with the motion of a reference video.
• tokenlab-talking-avatar - Portrait plus script or audio to a talking-head video.
• tokenlab-cinematic-video - Text or image to cinematic shots with camera language.

Audio - Voiceover, transcripts and subtitles, and music.

• tokenlab-voiceover - Script to natural speech audio with a chosen voice.
• tokenlab-transcribe-and-subtitle - Audio or video to transcripts, SRT/VTT subtitles, and translations.
• tokenlab-music - Songs, instrumentals, and sound effects from a brief.

3D and worlds - Downloadable 3D assets and explorable worlds.

• tokenlab-image-to-3d-asset - Image or text to a downloadable 3D model (GLB and more).
• tokenlab-world-generation - Text, image, or video to an explorable 3D world.

Decisions - Typed yes/no, choice, and score answers for agent branches.

• tokenlab-decisions - Typed yes/no, choice, and score answers with System One.

Integration - Put TokenLab into apps, native endpoints, cost routes, and OpenAI-compatible migrations.

• tokenlab-api-integration - Runnable TokenLab code for chat, native, media, and async tasks.
• tokenlab-native-endpoints - Use Responses, Messages, Gemini, and System One native routes.
• tokenlab-cost-routing - Cost-aware model routes and fallback chains from live pricing.
• tokenlab-openai-compatible-migration - Move OpenAI-compatible apps to TokenLab with a minimal diff.

Install

Claude Code plugin marketplace

/plugin marketplace add hedging8563/tokenlab-skills
/plugin install tokenlab@tokenlab-tools

Installed skill names are namespaced, for example /tokenlab:tokenlab-product-photography.

Codex plugin

Codex CLI 0.149.0 can consume this marketplace:

codex plugin marketplace add hedging8563/tokenlab-skills
codex plugin add tokenlab@tokenlab-tools

Skills CLI (any agent)

Install one skill at a time:

npx skills add https://github.com/hedging8563/tokenlab-skills --skill tokenlab-run -y
npx skills add https://github.com/hedging8563/tokenlab-skills --skill tokenlab-product-photography -y

Use any id from the catalog above.

Copy the folders

The skills follow the open Agent Skills specification. Clone the repository and copy the skills/<id>/ folders into the directory your agent loads skills from (for example .claude/skills/ or .agents/skills/).

git clone https://github.com/hedging8563/tokenlab-skills.git

API key and MCP servers

Media, audio, 3D, and decision calls need a TokenLab API key in TOKENLAB_API_KEY. Model discovery and pricing work without a key.

The bundled helpers send X-TokenLab-Client-Source so TokenLab can count which skills are used. tokenlab_run.py sends skills-<skill folder> when a skill runs it as ../tokenlab-run/scripts/tokenlab_run.py, and skills-run in other cases. search_api.py sends skills-tokenlab-api-integration. The header holds only the folder name.

The plugin bundles two optional MCP connections:

• tokenlab-stdio runs the published @tokenlabai/mcp-server over stdio. It uses the core tool profile by default, so outcome skills can call create_image, create_video, create_music, create_3d_model, create_speech, evaluate_decisions, get_task_status, and the other credentialed tools. Set TOKENLAB_API_KEY for those calls. Set TOKENLAB_MCP_TOOL_PROFILE=catalog for read-only model discovery, or full for world generation, webhooks, and batch tools. Restart Claude Code after you change it. The profile uses Claude Code's environment-variable default syntax. Changing the profile does not grant API permissions.
• tokenlab-hosted-model-explorer connects to the public hosted Model Explorer over Streamable HTTP. It is read-only and holds no API key.

The plugin and the Skills CLI are alternative ways to load the same skills. Neither changes the client's default model provider. To use TokenLab as a model provider, see the API integration skill's Codex, Claude Code, OpenCode, Pi, and Hermes setup references.

Update an existing installation

Release 0.2.0 adds the foundation skills, 21 outcome skills for image, video, audio, 3D, and decisions, catalog.json, and skills.sh.json. The stdio MCP server now defaults to the core profile instead of catalog. The five earlier skill ids are unchanged.

.claude-plugin/plugin.json owns the plugin version. A Git push alone does not refresh a cached Claude plugin whose version is unchanged; see Claude's version rules.

For a Claude Code plugin installation:

/plugin marketplace update tokenlab-tools
/plugin update tokenlab@tokenlab-tools

Restart Claude Code after updating. For a Codex plugin installation:

codex plugin marketplace upgrade tokenlab-tools
codex plugin add tokenlab@tokenlab-tools
codex plugin list --json

For skills installed with npx skills, update each skill in its original scope:

# In the project where you installed the skill:
npx skills update tokenlab-api-integration --project

# Or, if it was installed globally:
npx skills update tokenlab-api-integration --global

Repository rules

• Each folder in skills/ has a SKILL.md whose name equals the folder name.
• Outcome skills use the tokenlab- prefix so they never replace another vendor's skill with the same name.
• Exact model IDs appear only in the generated model block. To change a default, edit preferredModels in catalog.json and run python3 scripts/sync_models.py. Skills state no prices; the agent reads prices from the pricing API.
• python3 -m unittest discover -s tests checks the catalog, generated files, model blocks, versions, and helper scripts.