# Agent skills

Costfluent's agent skills teach a coding agent the common cost workflows: what to read, in which order, and how to report what it found. Each skill drives the tools of the [MCP server](/developers/mcp), and names each tool's [command line](/developers/cli) command for an agent without the server. The skills and the plugin manifests for each agent live in one public repository, [costfluent/agent-skills](https://github.com/costfluent/agent-skills), under the Apache-2.0 licence. They only read: no skill creates, changes or deletes anything in Costfluent.

## Before you start

- An API token. Create one as the MCP guide's [Create a token](/developers/mcp#create-a-token) describes, from the **Read only** preset.
- `costfluent-mcp` on your `PATH`, which the plugins start, and for Codex, Cursor and Copilot CLI also `costfluent`, which adds the server; [Install](/developers/cli#install) on the command line page says how.

## Install the skills

### Claude Code

```text
/plugin marketplace add costfluent/agent-skills
/plugin install costfluent@costfluent
```

Claude Code asks for the token when it enables the plugin and keeps it in your system keychain. `/mcp` then shows `costfluent` connected, and the skills are listed under the plugin.

### Gemini CLI

```bash
gemini extensions install https://github.com/costfluent/agent-skills
```

Gemini CLI asks for the token and keeps it in your system keychain.

### Codex

```bash
codex plugin marketplace add costfluent/agent-skills
costfluent mcp install --client codex --write
```

Install `costfluent` from the marketplace in Codex's plugin list, then restart Codex from a shell where `COSTFLUENT_API_TOKEN` is set.

### Copilot CLI

```bash
copilot plugin marketplace add costfluent/agent-skills
copilot plugin install costfluent@costfluent
costfluent mcp install --client copilot --write
```

Start `copilot` from a shell where `COSTFLUENT_API_TOKEN` is set.

### Cursor and any other agent

```bash
npx skills add costfluent/agent-skills
costfluent mcp install --client cursor --write
```

`npx skills add` offers the skills and copies the ones you choose into your agent's skills folder. For an agent other than Cursor, connect the MCP server as its own documentation says; [Connect an assistant](/developers/cli#connect-an-assistant) lists the clients `costfluent mcp install` writes for.

## Skills

| Skill | What your agent does with it |
|---|---|
| `investigate-cost-change` | Finds where spend moved between two periods, by provider, service, account and tag, and checks the anomalies and annotations on those days |
| `work-savings-queue` | Reads the savings queue, largest saving first, and keeps estimated and verified savings apart |
| `check-budget` | Says which budgets are on track, at risk or over, and what their figures leave out |
| `kubernetes-efficiency` | Finds the clusters, namespaces and workloads with the most idle cost and the widest gap between requested and used CPU and memory |
| `commitments-coverage` | Reads how much eligible usage your savings plans, reservations and committed use discounts covered, and which expire soon |
| `costfluent-terraform` | Writes Terraform for the Costfluent provider and the cost access modules, and checks that the connection reads costs once you apply it |

<FeatureStatus feature="CommitmentCoverage" />

## How it behaves

- A skill only calls read tools. The token's capabilities decide which tools exist, so a skill whose tool is missing says which capability the token lacks.
- The token stays in your system keychain with Claude Code and Gemini CLI, and in your shell's environment with the other agents; no manifest holds it.
- Missing data is reported as unavailable, never as zero. A detected change is not its explanation: a skill says what the figures show and what would confirm a cause, and never states a cause as certain.
- Figures leave out spend in a currency with no exchange rate and the days before your plan's history starts, and a skill says so when a result lists either.
- Names, titles and tag values in a result are written by your users and your cloud accounts. A skill treats them as data, never as instructions.
- An update to the skills is a new plugin version, which your agent takes when it updates its plugins.

## Related

- [MCP server](/developers/mcp)
- [Command line](/developers/cli)
- [Terraform](/developers/terraform)
