MCP/Asitis Agent Tools
Asitis Agent Tools logo

Asitis Agent Tools

Read-only linter for CLAUDE.md, AGENTS.md, SKILL.md, and rules files. Detects hidden instructions, invisible characters, token bloat, and drift.

developer-tools•Testing & Automation
GitHub12
License:MIT
VERIFIED
ByKatta@Katta

About this Server

Asitis Agent Tools is an offline, read-only linter and token context budget analyzer for agent steering files (CLAUDE.md, AGENTS.md, SKILL.md, .cursorrules). It catches prompt injections disguised as hidden HTML comments, zero-width spaces and bidi characters, shadowed instruction files, duplicate/contradictory rules, and rule length violations across Claude Code, Cursor, Codex, and Windsurf.

Real-time Interaction

Direct low-latency RPC protocol communication with structured JSON schemas.

Scoped Sandboxing

Restricted capability bounds to prevent accidental modifications or leakage.

Security & Best Practices

Strict Read-Only Architecture: The MCP server cannot modify files, run child processes, or make network calls.
Root Sandboxing: Scopes file reads strictly to the user-specified project root directory path.
Zero Telemetry: Operates 100% locally with zero external telemetry or cloud uploads.

When to Use (and When Not To)

Best For
  • •Auditing repositories for prompt injection vectors in documentation, pull requests, and forks
  • •Measuring and shrinking token context consumption in AGENTS.md, CLAUDE.md, and skill files
  • •Verifying SKILL.md frontmatter compliance with the Agent Skills specification
Not For / Limitations
  • •Web/mobile UI journey and regression testing (use agent-qa instead)
  • •Apple Developer documentation lookups (use Apple Developer Docs MCP instead)
  • •Database schema branching and SQL query execution (use DBHub, Neon, or PlanetScale MCP instead)

Install & Setup

terminal command
$claude mcp add asitis -- npx -y @asitis/mcp@0.3.0 --root "$PWD"

Exposed Tools & Endpoints

6 tools

Functions your AI agent can invoke via this Model Context Protocol server

lint_steering

Lints CLAUDE.md, AGENTS.md, and SKILL.md files for invisible unicode, prompt injection comments, duplicate rules, and vendor limits.

context_budget

Calculates approximate tokens loaded per session across Claude Code, Codex, Cursor, Gemini CLI, and Windsurf.

read_markdown

Reads markdown files with byte profiles, character counts, line endings, and security summaries.

find_steering_files

Lists all steering and rule files grouped by session-load timing (always, on-match, or on-invocation).

outline

Extracts ATX/setext headings and structure from markdown files while skipping code blocks.

suggest_fixes

Produces a unified diff preview of mechanical fixes for formatting errors and invisible characters.