> ## Documentation Index
> Fetch the complete documentation index at: https://docs.fianza.space/llms.txt
> Use this file to discover all available pages before exploring further.

# Claude Code skill

> Install the Fianza skill so Claude Code writes agent credit code against the real SDK surface instead of guessing at it.

Fianza ships a [Claude Code](https://claude.com/claude-code) skill. Install it
once and Claude writes agent-credit code against the real SDK — correct method
names, correct lifecycle order, and the failure modes that actually bite —
instead of inferring an API from the package name.

It covers both SDKs: `@fianza/agent-sdk` (JavaScript/TypeScript) and
`fianza-agent-sdk` (Python).

## Install

```bash theme={null}
npx @fianza/skill
```

Then restart Claude Code (or `/reload`) and invoke it with `/fianza-agent-sdk`.

The installer writes to `~/.claude/skills/fianza-agent-sdk/`. Re-running it
updates in place.

<Accordion title="Other install methods">
  **curl** — no npm required:

  ```bash theme={null}
  mkdir -p ~/.claude/skills/fianza-agent-sdk && \
    curl -fsSL https://raw.githubusercontent.com/TechnicallyKiller/Fianza/main/.claude/skills/fianza-agent-sdk/SKILL.md \
    -o ~/.claude/skills/fianza-agent-sdk/SKILL.md
  ```

  **Plugin marketplace:**

  ```
  /plugin marketplace add TechnicallyKiller/Fianza
  /plugin install fianza-agent-sdk@fianza
  ```
</Accordion>

## What it knows

The skill is not a link to these docs — it is a working reference Claude reads
before writing code.

| Area          | What the skill carries                                                                                                |
| ------------- | --------------------------------------------------------------------------------------------------------------------- |
| Method map    | Every SDK method in both languages, side by side, so Claude picks `creditLine()` vs `credit_line()` correctly         |
| Lifecycle     | `register` → `underwrite` → `creditLine` → `borrow` → `repay`, in the order the contracts require                     |
| Draw-on-402   | How `payWithCredit` auto-borrows a shortfall and settles an x402 payment                                              |
| Credit ramp   | Why `vaultState().limitUsdc` (enforced) differs from `creditLine().limitUsdc` (scored), and how repaying grows it     |
| Networks      | `network: "mainnet"` vs the testnet default, and which calls have no mainnet equivalent yet                           |
| Errors        | `ValidationError`, `ApiError`, `TxError`, `MaxDrawExceededError` — which are thrown locally and which come from chain |
| Real pitfalls | SAC error #10 and #13, why a fresh agent correctly scores 0, backend cold starts                                      |

That last row is the point. A fresh agent scoring `400 / Unrated / limit 0` is
the honest result of having no revenue, not a bug — but an assistant without the
skill will usually treat it as one and start debugging the wrong thing.

## Example

With the skill installed, this prompt is enough:

> Write a Python agent that registers on Fianza, gets underwritten on its
> on-chain revenue, then buys a \$0.05 x402 research call on credit.

Claude will reach for `pay_with_credit`, pass `skip_proof=True` to avoid the
slow zkTLS proof, and handle `MaxDrawExceededError` — without being told any of
that.

## Keeping it current

The skill is versioned with the SDK surface it documents. `npx @fianza/skill`
again to update; it overwrites in place.

<Info>
  The earlier `@trustline-agents/skill` package still works and installs the same
  skill under its former name. New installs should use `@fianza/skill`.
</Info>

## Related

* [SDK reference (JavaScript/TypeScript)](/sdk-reference)
* [SDK reference (Python)](/sdk-reference-python)
* [Onboarding kit](/onboarding-kit) — getting a funded testnet agent from zero