Skip to content

liquid4all/liquid-d1

v0.1.0MIT

Build typed System One decisions with Liquid AI's d1 family: hosted d1 and d1-ultra, plus local d1-3B and d1-omni-600M with image and audio support.

Liquid d1 agent skill

Drop-in skill for Claude Code, Codex, and other agent environments.

The skill gives a coding agent context on Liquid AI's d1 decision-model family: the shared System One contract (Choice, Score, Noul), question design, probabilities, hosted API integrations, and local inference.

Choose a d1 model

ModelDeploymentInputs
d1Hosted; lower-priced optionText, images
d1-ultraHosted; most capable optionText, images
d1-3BOpen weights, 3.1BText, images
d1-omni-600MOpen weights, 587MText with images or audio

See the model overview and migration guide.

Access d1

  • Liquid direct: d1 or d1-ultra at https://api.liquid.ai/decisions/v1/systemone
  • Vercel AI Gateway: liquid/d1 or liquid/d1-ultra through AI SDK's decision API
  • OpenRouter: liquid/d1 or liquid/d1-ultra through the Decisions API
  • Local / self-hosted: d1-3B or d1-omni-600M through Transformers, llama.cpp, or MLX; see local inference.

The dependency-free request runner supports hosted Liquid and compatible local HTTP servers. For a local llama.cpp server, pass a request file containing state and questions (no hosted model ID):

LIQUID_BASE_URL=http://127.0.0.1:8080 LIQUID_API_KEY= \
  python3 skills/liquid-d1/scripts/d1.py request.json

Local servers can omit authentication. Set LIQUID_API_KEY if your server requires a bearer token. The no-argument demo uses hosted model: "d1"; --models requires a server implementing GET /v1/models.

Installation

Claude Code

Once this is published to the Liquid4All/liquid-d1-skills marketplace repo:

claude plugin marketplace add Liquid4All/liquid-d1-skills
claude plugin install liquid-d1@liquid-ai

Codex

Add the repository as a marketplace source:

codex plugin marketplace add Liquid4All/liquid-d1-skills
codex plugin add liquid-d1@liquid-ai

Alternatively, open /plugins after adding the marketplace and install Liquid d1 there. Start a new chat before using it.

Other agents

npx skills add Liquid4All/liquid-d1-skills --skill liquid-d1

Choose your agent when prompted. Installation is project-local by default; add -g to install globally.

Manual

Copy the entire skills/liquid-d1/ directory, including references/ and scripts/, into your agent's skills directory (for Claude Code: .claude/skills/liquid-d1/ in the project, or ~/.claude/skills/liquid-d1/ globally).

Paste into your agent

Install the Liquid d1 skill. If you're in Claude Code, run
`claude plugin marketplace add Liquid4All/liquid-d1-skills`, then
`claude plugin install liquid-d1@liquid-ai`. If you're in Codex, run
`codex plugin marketplace add Liquid4All/liquid-d1-skills`, then
`codex plugin add liquid-d1@liquid-ai`. If you're in another agent, run
`npx skills add Liquid4All/liquid-d1-skills --skill liquid-d1` and select your agent.
Use one installation method. Then use the Liquid d1 skill when working on
this project.

Choose one installation method to avoid duplicate copies.

Updates

Claude Code plugin:

claude plugin marketplace update liquid-ai
claude plugin update liquid-d1@liquid-ai

Codex marketplace:

codex plugin marketplace upgrade liquid-ai

Restart Claude Code or run /reload-plugins. For skills.sh installs, run npx skills update. For manual copies, replace the whole directory.

Example prompts

Naming the skill in the prompt works in any agent. With the Claude Code plugin you can invoke /liquid-d1:liquid-d1; in Codex, use $liquid-d1.

Find where d1 fits in an existing codebase:

Using the Liquid d1 skill, explore this project and find places where a
typed decision from d1 could replace fragile parsing, regex classification,
or a prompt-and-parse LLM call.

Let the agent experiment against the API. Create a key on the console's API keys page and export it:

Using the Liquid d1 skill, run some experiments with the key I've exported
to LIQUID_API_KEY. Propose changes based on the most promising results.

Working with the agent

  1. Talk it through first. Use the prompts above as a starting point.
  2. Review the plan before implementing.
  3. Keep questions and thresholds in one file. They are what humans most need to review, and agents are not great at writing questions on the first pass. Expect to edit them together.
  4. Don't take the agent's assertions at face value. Ask it to run the representative cases and show the probabilities.

Common issues

The agent isn't using the skill. In Claude Code, invoke /liquid-d1:liquid-d1; in Codex, invoke $liquid-d1. In other agents, say "use the Liquid d1 skill". If it still does not load, confirm the installer targeted the agent you are using, then restart it.

Routing isn't doing what you expect. Check the questions and thresholds. Thresholds set too high cause false negatives; too low, false positives. The questions may also need to be more specific about what the criteria mean.

Confidence thresholds everywhere. If you only need the best option, take the highest-probability choice. If you have a specific statistical rule in mind, use the probabilities directly. Confidence describes how concentrated the distribution is, nothing more.

The agent invents request or response fields. The skill is stale, or the agent didn't read the live docs. Update the skill, and point it at https://console.liquid.ai/docs.

401 on every request. Check that LIQUID_API_KEY is set in the environment the agent is running in and that the key is valid.

429 with model_unavailable. No healthy inference target for the model right now. This is not a quota problem; retry after a short wait.

Files

liquid-d1-skills/
├── .agents/plugins/marketplace.json  Codex marketplace catalog
├── .claude-plugin/                   Claude plugin and marketplace manifests
├── plugin.json                       portable Agent Plugins manifest
└── skills/liquid-d1/
    ├── SKILL.md                      instructions the agent loads
    ├── references/
    │   ├── api.md                    API contract, limits, and errors
    │   ├── providers.md              Liquid, Vercel, and OpenRouter access
    │   ├── sdks.md                   curl, Python, and TypeScript usage
    │   └── local-inference.md        open weights, images, and audio
    └── scripts/
        └── d1.py                     stdlib-only request runner

License

Skill code and documentation: MIT © 2026 Liquid AI.

The downloadable d1-3B and d1-omni-600M model weights use the LFM Open License.