claude-api-toolkit
An Agent Skill for writing, reviewing, debugging and migrating code that calls the Claude API.
What makes it different: its model tables and its request linter were built by sending 46 request
shapes to each of 15 Claude models and recording what the API accepted (see
references/request-rules.md). The linter's verdict matched the API on every recorded outcome.
Not affiliated with Anthropic. "Claude" is a trademark of Anthropic, PBC.
Contents
| Path | Purpose |
|---|---|
SKILL.md |
When to use the skill, the working loop, the rules that cause most failures |
references/models.md |
Ids, context, output caps, prices (generated) |
references/request-rules.md |
What each model accepts, with exact error texts (generated) |
references/requests.md |
Request anatomy, streaming, structured outputs, files, batches, token counting |
references/tools-and-agents.md |
Tool definitions, the tool loop, server tools, agent architectures |
references/managed-agents.md |
Core Managed Agents flow |
references/caching-and-context.md |
Prompt caching, compaction, context editing, cost levers |
references/errors.md |
Status codes and observed 400 messages with fixes |
scripts/claude-api.mjs |
Offline CLI: models, lint, fix, migrate, cost, compare, scaffold, guide |
The scripts use only the Node.js standard library, open no network connection and need no key.
Install
skills install @spboucher/claude-api-toolkit --client claude-code # or codex, cursor, generic-agent-skillsMeasured effect
tests/skill-eval in the source repository asks a model to write Messages requests for specific
models, with and without this skill in its system prompt, and sends each result to the live API.
The measured acceptance rates are in CHANGELOG.md. It is a small evaluation (see the numbers and
their caveats there), not a benchmark.
Limits
A snapshot taken on the date in SKILL.md under API version 2026-10-21. Models the probing key
could not see are absent. Computer use, the advisor tool, memory stores, message threads and
cloud-platform differences were not probed. Cost output is an estimate. When the API and this
skill disagree, the API is right.