Agent Skills

Teach your coding agent Confident AI workflows with official Agent Skills.

Overview

Confident AI publishes official Agent Skills that teach your coding agent — Cursor, Claude Code, Codex, Windsurf, or any other Skills-compatible assistant — how to work with Confident AI correctly. A skill is a SKILL.md file plus reference docs and code templates that your agent reads before it acts, so it follows the same workflows our own docs prescribe instead of guessing from training data.

GitHub Repository

View the source code, report issues, and find the latest releases.

Where the MCP server gives your agent live access to your project's resources (traces, datasets, prompts, dashboards), skills give it knowledge — how to set up evals, instrument an app, or provision a project the right way. Most teams use both together, and the confident-eval-setup skill works entirely through the MCP server.

Official Skills

There are six official skills, each with a deliberately narrow scope so your agent picks the right one automatically:

Skill	What it teaches	API key it uses
`confident-client`	Work with the Confident AI API through the Admin SDK — account administration plus every project resource	`CONFIDENT_ORG_API_KEY` and `CONFIDENT_PROJ_API_KEY`
`confident-eval-setup`	Connect your app as an AI connection over the MCP server, verify it with a ping, and upload a dataset to evaluate	None (MCP server, OAuth)
`confident-agent-handler`	Connect an app with no HTTP endpoint through the Confident Agent's handler mode, then create and ping its AI connection	`CONFIDENT_API_KEY` (run by the Confident Agent)
`confident-tracing`	Instrument Python and TypeScript AI apps with `confident-trace` integrations and custom spans	`CONFIDENT_API_KEY`
`confident-otel`	Export raw OpenTelemetry AI traces to Confident AI from any language without the `confident-trace` package	`CONFIDENT_API_KEY`
`confident-migrate-langfuse`	Move tracing from Langfuse SDK v1–v4 to `confident-trace`, or export Langfuse's spans over OpenTelemetry	`CONFIDENT_API_KEY`

The scopes are mutually exclusive by design. Each skill's trigger description tells the agent when not to fire and which sibling skill to use instead, so "set up evals for my agent" activates confident-eval-setup, "connect our pipeline that has no API" activates confident-agent-handler, "add tracing" activates confident-tracing, "export our existing OTel spans" activates confident-otel, "migrate us off Langfuse" activates confident-migrate-langfuse, and "create a project for the new team" activates confident-client — without you naming a skill.

Where the Skills Live

Every skill lives in the confident-ai/confident-skills repository, grouped into plugins:

The confident-eval-setup and confident-agent-handler skills are the confident-eval-setup plugin.
The confident-tracing, confident-otel, and confident-migrate-langfuse skills are the confident-trace plugin. They teach the Python and TypeScript SDKs in confident-ai/confident-trace.
The confident-client skill is the confident-client plugin. It teaches the Python and TypeScript Admin SDKs in confident-ai/confident-client.

Installation

Every skill installs the same three ways. Each skill's own page lists its exact commands; the shapes are:

Works with Cursor, Claude Code, Codex, Windsurf, OpenCode, and any other assistant that supports the Skills standard:

npx skills add confident-ai/confident-skills --skill "confident-eval-setup"

Swap the skill name for the one you want — for example --skill "confident-tracing".

The confident-skills repository doubles as a Claude Code plugin marketplace. To install the tracing skills:

/plugin marketplace add confident-ai/confident-skills
/plugin install confident-trace@confident-skills
/reload-plugins

The plugin bundles confident-tracing, confident-otel, and confident-migrate-langfuse. See Plugins for the evals, administration, Cursor, and Codex commands.

Copy or symlink the skill folder into your agent's skills directory:

git clone https://github.com/confident-ai/confident-skills
cp -r confident-skills/plugins/confident-trace/skills/confident-tracing .claude/skills/

For Claude.ai on the web, zip the skill folder and upload it under Settings → Capabilities → Skills.

Once installed, you don't invoke a skill explicitly — describe what you want and the agent activates the matching skill on its own.

FAQs

Do I need the MCP server if I install the skills?

They solve different problems and work best together. The MCP server gives your agent live access to your project's resources — traces, datasets, prompts, dashboards. Skills teach it the workflows: how to set up evals, instrument an app, or provision a project correctly. The confident-eval-setup skill needs the MCP server, since every step it takes is an MCP tool call, and confident-agent-handler uses it to create and ping the connection; the other skills work without it.

Which coding agents support skills?

Any assistant that implements the Skills standard — Cursor, Claude Code, Codex, Windsurf, OpenCode, and others. Claude.ai on the web also accepts skills as zipped uploads under Settings → Capabilities → Skills.

How do the skills stay up to date?

A Skills CLI install or manual copy is a static snapshot — rerun the install to pick up changes. If you install through a plugin instead, your client tracks the repo as a marketplace source and updates through its own plugin flow.

Are the skills open source?

Yes — all six are Apache-2.0 licensed and live in the public confident-ai/confident-skills repository, so you can read every instruction a skill gives your agent before installing it.

Next Steps

MCP Server

Give your agent live access to your project's traces, datasets, prompts, and evals — skills pair naturally with it.

Custom Agent Skills

Serve your own project-specific onboarding skills to coding agents from Confident AI's git endpoint.

Scaling beyond prototype?For teams evaluating Confident AI in productionTalk to us

Last updated on November 6, 2026