calcifier

Calcifier turns Claude/Codex skills into executable Python workflows. Give it a SKILL.md; it produces a recipe that you can read, review, run, and debug. Python handles repeatable operations. Agents handle reasoning.

Why use it?

A skill is a set of instructions that an agent interprets each time it runs. That works well for exploration, but a recurring task often has a stable procedure: gather files, run a script, ask for a judgment, then write a report. Repeating that planning costs model calls and makes the procedure harder to inspect.

Calcifier makes the procedure explicit. For example, a review skill can become Python that selects files, calls an agent to assess them, validates the returned findings, and writes the report. You can change the file selection or report format without asking a model to invent the whole workflow again.

Use it for skills you run repeatedly, such as code reviews, data triage, and report generation, when you want:

• A workflow you own: ordinary Python functions and control flow that humans can maintain.
• Less repeated model work: file operations, scripts, transformations, and rendering execute directly in Python.
• Checked answers: agent responses are parsed into typed values with explicit validation and a bounded retry.
• Visible execution: budgets, permissions, artifacts, and traces make it easier to see what ran and where it failed.

Agent answers remain nondeterministic. Compilation checks catch mechanical errors; whether a recipe follows the skill and produces good answers needs testing on real tasks. Recipes execute Python on your host, so review generated code before granting trust.

Try it

Install from source until the first PyPI release. Requires Python 3.13+, uv, and an authenticated Claude or Codex CLI (or a custom agent adapter).

uv tool install git+https://github.com/trailofbits/calcifier.git
calcifier --version

Follow the quickstart to configure agent roles, compile a skill, review the generated recipe, and run it.

Compilation starts with a runnable baseline that gives one agent session the intact skill and its resources. It then tries to move repeatable work into Python. If that fails, it retains and labels the baseline; use --require-compiled when you need compilation failure to exit nonzero.

The Claude Code plugin provides compilation, execution, debugging, and update skills:

/plugin marketplace add trailofbits/calcifier
/plugin install calcifier@calcifier

For local testing, use the checkout's absolute path as the marketplace.

Commands and documentation

Use calcifier COMMAND --help for flags.

• Usage: setup, compilation, inputs, trust, outputs, and traces.
• Recipe authoring: typed answers, effects, budgets, parallel work, and composition.
• Providers: agent roles and custom agent/tool contracts.
• Architecture: compiler and runtime internals.
• Contributing: development, testing, and compiler replay.

Development

uv --no-config sync --all-groups --locked
gmake verify

See CONTRIBUTING.md and SUPPORT.md for contribution and support policies. Report vulnerabilities privately using SECURITY.md. Project interactions follow the Code of Conduct.

MIT, copyright Trail of Bits, Inc.