Marvin is a Python framework for producing structured data with language models and coordinating tasks carried out by specialized AI agents. Developers use it to extract, cast, classify, or generate data and to compose observable agentic workflows. The catalogue includes an instruction and an agent for working with Marvin.
Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/prefecthq/marvin/claude-mdgit clone --depth 1 https://github.com/PrefectHQ/marvinWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/prefecthq/marvin/claude-md)<a href="https://agentmods.dev/instructions/prefecthq/marvin/claude-md"><img src="https://agentmods.dev/badge/instructions/prefecthq/marvin/claude-md.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00462 | $0.00462 |
| Opus 5 | $0.00231 | $0.00231 |
| Sonnet 5 | $0.00092 | $0.00092 |
| Haiku 4.5 | $0.00046 | $0.00046 |
Grade A, and why
marvin CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 6d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
Marvin - AI Engineering Toolkit
Marvin is a lightweight AI engineering toolkit for building natural language interfaces that are reliable, scalable, and easy to trust.
Reproductions
- use the repros folder to reproduce the results (e.g.
uv run repros/1234.py) - this folder is not checked into git
Architecture & Design Philosophy
- Aggressively minimal and elegant: Keep implementations simple and focused
- Functional first: Prefer functional approaches, use classes where justified
- Type-safe: Full type annotations, modern Python syntax (3.10+)
- Private internals: Keep implementation details "private" (e.g.
def _impl)
Key Components
- Engine: Core AI interaction layer
- Tasks: Structured AI task definitions and execution
- Tools: Extensible function calling capabilities
- Agents: AI agents with tool access and memory
- Memory: Persistent conversation and context storage
- Handlers: Event processing and routing
- CLI: Command-line interface for common operations
Development Guidelines
Type Hints
- Use
X | Yinstead ofUnion[X, Y] - Use builtins like
list,dictinstead oftyping.List,typing.Dict - Use
T | Noneinstead ofOptional
Dependencies & Running
- Use
uvfor dependency management and script execution - Install deps:
uv syncoruv sync --extra foo - Run scripts:
uv run some/script.pyoruv run --with pandas script.py - Testing:
uv run pytestoruv run pytest -n3for parallel
Finding Things
- Use
rgfor searching, not grep - Use
lsandtreefor navigation - Check git context with using the GitHub MCP server
- Think like a hacker with good intentions - search in site-packages when needed
Linter Philosophy
- Empirically understand by running code
- Linter tells basic truths but may be orthogonal to goals
- Don't obsess over upstream linter errors, use as clues when relevant
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 6d ago First seen · 48 lines · 462 tokens per session scan A 6be3ef99cdcf
marvin CLAUDE.md is an instructions file published in the GitHub repository PrefectHQ/marvin (6,194 stars, last pushed 15d ago), licensed Apache-2.0. It adds 462 tokens to every session, about $0.0023 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
langwatch CLAUDE.md
Claude Code instructions for langwatch/langwatch, covering langwatch, before you implement anything, development environment, running with no container runtime and local dev by hostname — thuishaven / portless (recommended).
dify AGENTS.md
AGENTS.md instructions for langgenius/dify, covering agents.md, repository gotchas and frontend workflow.
skyvern CLAUDE.md
Claude Code instructions for Skyvern-AI/skyvern, covering claude.md, development commands, python backend commands, code quality & testing and frontend commands (in skyvern-frontend/).
skyvern AGENTS.md
AGENTS.md instructions for Skyvern-AI/skyvern, covering skyvern agent guide, project structure for agent navigation, coding conventions for agents, python standards and asynchronous programming.
OllamaSharp AGENTS.md
Instructions for awaescher/OllamaSharp, covering agents.md — ollamasharp, project overview, how ollamasharp fits in the .net ai ecosystem, why this matters especially for tool use / function calling and build / test / lint commands.
dify CLAUDE.md
Claude Code instructions for langgenius/dify, a project described as: Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace. Deploy on cloud, VPC, or self-hosted, so teams move from prototype to production without rebuilding the stack.