mithril CLAUDE.md

A Rust-based sandboxed MCP server that gives AI agents validated execution tools instead of unrestricted shell access.

In plain words
What is it for?
Use it to build or test the server and its tools for Git, Docker, Go, JavaScript, Python, Rust, and related workflows.
Why use it?
It confines commands through typed inputs, injection checks, path restrictions, and operating-system isolation.

Instructions file

Install

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.

agentmods
npx agentmods add instructions/radimsem/mithril/claude-md
Clone the repo
git clone --depth 1 https://github.com/radimsem/mithril
Per session 1,709 This file is loaded in full into every session.
When invoked 1,709 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5 $0.01709 $0.01709
Opus 5 $0.00855 $0.00855
Sonnet 5 $0.00342 $0.00342
Haiku 4.5 $0.00171 $0.00171

Measured 3d ago against content hash 83b6a6245896, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

mithril 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 3d 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.

CLAUDE.md · 135 lines

How it starts

The opening of the file, as written. The whole thing — 135 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Mithril

Sandboxed MCP server that replaces the generic shell tool with validated, purpose-built execution tools for AI agents. No shell, no escapes — every command goes through typed arguments, injection guards, path confinement, and OS-level isolation.

Quick Reference

cargo build                        # Build all crates
cargo build -p mithril             # Binary only
cargo test                         # All non-ignored tests
cargo test -- --ignored            # Include sandbox + ecosystem tests
cargo test -p server               # Single crate
cargo test -p server cat_reads_file # Single test by name
cargo clippy --workspace           # Lint
cargo fmt --all -- --check         # Format check

Workspace Structure

crates/
  bin/        CLI binary entry point (clap args, config loading, MCP transport)
  common/     Shared types: Config, ExecContext, ToolDef/ToolExec traits, errors, validation rules
  patterns/   Compiled regex patterns for validation (melody DSL, lazy_static)
  tools/      ~160 tool implementations across 10 domains (system, git, docker, go, js, python, rust, gh, glab, run_chain)
  server/     MCP ServerHandler, tool/prompt/resource routing, execution pipeline, history buffer
  runner/     Process spawning layer: timeouts, stdout streaming, progress notifications
  validator/  Argument validation layer: injection guard, path confinement, env blocklist, elicitation
  sandbox/    OS isolation layer: BubbleWrap (Linux) / sandbox-exec (macOS) / Noop fallback
  proxy/      Optional RTK (Rust Token Killer) prefix layer for token-optimized output

Architecture

Execution Pipeline

Every tool call flows through 4 sequential layers:

MCP Request → Router → build_context() → [ Validate → Proxy → Sandbox → Runner ] → format() → MCP Response
  1. ValidateLayer — checks argument types, string patterns, injection characters, path confinement, env blocklist, elicitation approval
  2. ProxyLayer — optionally prepends rtk binary for supported commands
  3. SandboxLayer — prepends OS-specific sandbox command prefix (bwrap/sandbox-exec)
  4. RunnerLayer — spawns process via tokio::process::Command, streams stdout, enforces timeout

Read the full file on GitHub · 135 lines

Changes

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.

  1. 3d ago First seen · 135 lines · 1,709 tokens per session scan A 83b6a6245896

Subscribe to this mod's changes

mithril CLAUDE.md is an instructions file published in the GitHub repository radimsem/mithril (16 stars, last pushed 4mo ago), licensed Apache-2.0. It adds 1,709 tokens to every session, about $0.0085 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.