Guardrails for the new AI coding era
Mustel is a local, non-AI static analysis layer that intercepts editor save loops. It formats and compresses code diagnostics into ultra-short prompts to guide coding agents without context bloat.
pip install mustel
Your AI IDE wastes piles of tokens
Standard coding agents transfer thousands of lines of raw, duplicate linter logs on every save loop. These verbose payloads clog context windows, increase latency, and inflate your API bills.
Mustel supports a huge ecosystem
Out-of-the-box rule configurations covering 26 core packages, cloud SDKs, web frameworks, and system utilities.
Deterministic Code Quality
Mustel executes linter checks, deduplicates overlap reports, and runs custom YAML pattern checks locally. It formats all issues and creates a sub-200 character summary to instruct AI agents clearly.
Fast file stat validation (mtime + size) returns findings in under 30ms.
Written-in-Rust linter tools (Ruff, Oxlint) scan scripts concurrently.
Mustel MCP Runtime Pipeline
mustel map [PATH]
Compiles files into structured AST outline nodes, reducing prompt context overhead by 95%.
System Benchmarks
Measured empirically via tiktoken (`cl100k_base` and `o200k_base` encodings) on real codebases.
Context Token Savings
Empirical reduction in total input tokens sent to LLMs.
Output Token Reduction
By presenting pre-summarized errors, LLMs write solutions directly.
Independent Recall
Measured against external, un-tuned test code. 100% on customized benchmarks.
Benchmark Test Subjects
| Project Name | Source | Files | Lines | Vulnerabilities | Detection Rate |
|---|---|---|---|---|---|
| project_auth | Internal | 1 | 82 | 5 | 100% |
| project_backend | Internal | 1 | 83 | 5 | 100% |
| project_independent | Un-tuned | 2 | 192 | 16 | 75% |
| EVFA | Open-source | 23 | 453 | 9 | 55.6% |
How Mustel Integrates
Mustel runs locally and hooks directly into your editor or terminal pipeline.
Local Scanning
Run `mustel scan` or initialize the MCP server. Mustel runs Ruff linting, Bandit security audits, and pip-audit CVE lookups locally in milliseconds.
Pattern Matching
It parses your python files and runs YAML-defined regex rules to scan for anti-patterns across 22 popular libraries.
Structured Context
An MCP server outputs the compressed findings into the system instructions, allowing the coding agent to query issues by ID and solve them immediately.