ESLint + Jest + npm publish — for AI Agent Tools.

Build, test, scan, and ship tools across OpenClaw/ClawHub, Claude Code, Cursor, and Windsurf — from a single CLI.

13,000+ tools are published on ClawHub. 13% contain critical security flaws (Snyk ToxicTools Report, Feb 2026). Tools break silently on platforms other than the one they were tested on. There is no pytest for agent tools — until now.

toolmark init my-tool --template github-api
toolmark test          # LLM-as-judge evaluation
toolmark scan          # prompt injection, dynamic fetch, credential leaks
toolmark compat        # check all 4 platforms at once
toolmark publish       # sign with Ed25519, push to ClawHub + Claude Code

pip install toolmark

Requires Python 3.12+.

# 1. Scaffold
toolmark init my-github-tool --template github-api

# 2. Edit tool.md and tests/
cd my-github-tool

# 3. Test
ANTHROPIC_API_KEY=sk-ant-... toolmark test

# 4. Scan
toolmark scan

# 5. Check platform compatibility
toolmark compat

# 6. Publish
toolmark publish --platforms clawhub,claude-code

toolmark init my-tool --template github-api      # GitHub REST API wrapper
toolmark init my-tool --template file-ops         # Local filesystem tool
toolmark init my-tool --template mcp-integration  # Wraps an MCP server tool
toolmark init my-tool --template web-search       # Search API tool
toolmark init my-tool --template loom-query       # Loom knowledge graph tool
toolmark init my-tool --template blank            # Minimal scaffold

# tests/test_search.yaml
- id: search_open_prs
  input: "find my open pull requests"
  expect_invoked: true
  expect_tool: search_pull_requests
  expect_params:
    state: open
    assignee: "@me"
  tolerance: fuzzy     # strict | fuzzy | invoked
  tags: [smoke]

Run: toolmark test --tags smoke

toolmark catches:

• SF001 — Dynamic fetch (curl | bash, eval(fetch(...)))
• SF002 — Hardcoded credentials (API keys, passwords)
• SF003 — Prompt injection phrases in tool descriptions
• SF004 — Undeclared network endpoints
• SNYK-* — 138 rules via Snyk agent-scan (if installed)

Every published tool is signed with Ed25519:

toolmark keygen              # creates ~/.toolmark/signing.key
toolmark publish --sign      # signs + publishes
toolmark verify my-tool     # verify any published tool

Every toolmark init project includes a ready-to-use workflow:

# .github/workflows/toolmark.yml — already in your project
- toolmark compat    # platform check
- toolmark scan      # security gate
- toolmark test      # LLM evaluation (needs ANTHROPIC_API_KEY secret)

See how your tool ranks: toolmark.dev/leaderboard

Quality Score = test pass rate (50%) + security score (30%) + compat score (20%).

• init — scaffold with 6 templates
• test — LLM-as-judge evaluation
• scan — built-in security rules + Snyk integration
• compat — 4-platform compatibility matrix
• bench — composite quality score
• publish — Ed25519 signing + ClawHub
• watch — re-run tests on save
• VS Code extension
• Rust benchmark runner
• Claude Code + Cursor + Windsurf publish

See CONTRIBUTING.md. We always have good first issues.

MIT — see LICENSE.

Built by @ddevilz as part of the Loom AI tooling ecosystem.