Skip to main content
A registry of agent skills with test results. See what a skill changes, then add it to the agent you already have.

​
Using skills

A skill is a SKILL.md file of instructions your agent reads when a task calls for them, such as how to handle a refund or how to use a particular tool. Your agent sees a short list of its skills and opens the one a task needs, so a skill costs a line of context until it is used.A skill is not a tool. A tool is a function the agent calls. A skill shapes how the agent works, and it can bundle scripts the agent may run. See How it works and Write a skill.
No. Without an account you can browse skills and their results, add one to Claude Code, Cursor or your own Python project, and run your agent. With no DecimalAI key, agent_config() sends nothing when your agent runs.A free account adds three things: change an agent’s prompt and skills without redeploying, see which skills each run used, and, on the OpenAI Agents SDK, test with and without a skill. You still bring your own model provider’s key, because DecimalAI never hosts your agent. See the quickstart.
Claude Code, Cursor and other editors read skill files from your project. In Python, eight frameworks have an agent_config() that gives your agent its skills: LangChain and LangGraph, the OpenAI Agents SDK, the Claude Agent SDK, the OpenAI SDK, the Anthropic Messages API, the Gemini API, Pydantic AI and Google ADK. CrewAI, Deep Agents, Microsoft Agent Framework and Strands load skills with their own loader. LlamaIndex and OpenTelemetry trace only.All frameworks shows which way works on each.
No. The Python SDK is the only first-party SDK. Skills are plain files, so any project can use them: see Cursor and other tools. The REST API works from any language (Authentication), and the decimalai-mcp server gives MCP clients read-only search.
Python 3.10 or later. The [adk], [pydantic-ai] and [all] extras need 3.11. On an older Python, pip quietly installs an older decimalai instead of failing, so put a floor on it: pip install "decimalai>=0.27". Init & Setup has the details.

​
Trusting and paying for skills

A skill’s page shows a benchmark: the same test cases run with the skill and without it, for one model and one version of the skill. It counts the cases the skill fixed, the ones that passed either way and the ones it broke, so read broke first.SkillScore (0 to 100) sums the evidence from verified benchmarks, live evaluations and AI ratings. A skill without enough evidence shows New. Results depend on the model and the cases, so treat a score as a shortlist and test a skill on your own cases. See How to read a skill page and SkillScore.
Every public skill is scanned before it is listed, and again for each new version, by a pattern scan, an AI review of intent and a content review. The result is one band: Passed, Caution, Blocked or Not yet reviewed. A Blocked skill can’t be installed or forked, and the runtime refuses to load it.A Passed band means nothing known tripped. It does not mean a skill is safe to run with your production credentials. Pin the version you reviewed, grant only the tools it needs, and read its scripts. See How skills are vetted.
Finding, reading and adding public skills is free, with no signup. A free account includes 2 agents, 10 skills of your own, 50 linked registry skills, 5,000 traces a month and 100 hosted benchmark cases a month. Core ($49 a month) raises those limits and adds publishing to the public registry. Pro is $199, and Enterprise is custom. See Pricing for every limit.Your model provider bills you for tokens, as it already does. A skill adds a line to your prompt, plus its full text when the model opens it.
Yes. A skill you create is not public: it starts visible to you only or to your organization, and becomes public only when you publish it, which runs the publish checks. Adding a public skill moves nothing of yours. To edit one, fork it: the fork is your own copy. See Update, pin or fork a skill.
Yes. Skills you add as files stay in your repo and keep working without DecimalAI. Your datasets, traces and organization’s skills export through the CLI and API, and deleting traces or a team is self-serve. The formats (SKILL.md, eval.yaml, manifests) are open, and the SDK, CLI and GitHub Action are open source. The hosted platform can’t be self-hosted today. See Trust & Exit.

​
Your data

An agent with a key sends each run’s trace: its name, the first 200 characters of the input and output, each model call’s messages and completion, tool calls, token counts, and which skills were offered and loaded. Text is stored as written, so scrub sensitive text first. Production runs on Google Cloud in us-central1.Security has the retention windows, encryption and every deletion path, including how to delete data.
Your agent calls your provider directly, with your key. Two kinds of feature are different. The Playground and fine-tuning use a provider key you store, encrypted. Server-side LLM judges and call replay in real mode use DecimalAI’s key and count against your plan. See Trust & Exit.

​
Regression checks

It compares a proposed change to your agent’s manifest (its tools, prompts, models and skills) against your recorded traces, and reports which runs depend on what changed. It needs no eval suite and doesn’t run your agent. Eval-driven testing runs cases you wrote against a new version. The two complement each other: see how they compare and Manifests.

​
Help

Free plans use the community on GitHub issues. Core has email support at hello@decimal.ai. Pro adds a Slack channel, and Enterprise has both. For a 5xx error, include the X-Request-ID response header so we can find the request in our logs.