moltblock — Trust Layer for AI Agents

Description

Moltblock provides verification gating for AI-generated artifacts. It runs policy checks and optional code verification (via vitest) to catch dangerous patterns before they reach production.

What moltblock does:

• •Generates code via LLM API calls, then runs policy checks against the output
• •When --test is provided, executes vitest to verify generated code against a user-provided test file (see Security: Test Execution below)
• •Reads its own config files (moltblock.json, ~/.moltblock/moltblock.json) if present
• •API keys are read from environment variables at runtime and sent only to the configured LLM provider endpoint

When to Use

Use moltblock when the task involves:

• •High risk (always verify): Destructive operations (rm -rf, DROP TABLE), privilege escalation (sudo, chmod 777), credential/key access (.ssh/, id_rsa, .env), system modification, raw disk writes
• •Medium risk (verify when uncertain): Network requests (curl, wget, fetch), file writes, database modifications, subprocess spawning, dynamic code evaluation
• •Low risk (skip verification): Simple text responses, math, reading public information, code that doesn't touch the filesystem or network

Tool: moltblock

Verify a task before execution.

Usage

npx moltblock@0.11.0 "<task description>" --provider <provider> --json

Parameters

Environment Variables

Moltblock auto-detects the LLM provider from whichever API key is set. If no key is set, it falls back to a local LLM at localhost:1234. Set one of these for a cloud provider:

• •OPENAI_API_KEY — OpenAI (primary)
• •ANTHROPIC_API_KEY — Anthropic/Claude (optional)
• •GOOGLE_API_KEY — Google/Gemini (optional)
• •ZAI_API_KEY — ZAI (optional)

Example

# Verify a task
npx moltblock@0.11.0 "implement a function that validates email addresses" --json

# Verify code with tests
npx moltblock@0.11.0 "implement a markdown-to-html converter" --test ./tests/markdown.test.ts --json

Output (JSON mode)

{
  "verification_passed": true,
  "verification_evidence": "All policy rules passed.",
  "authoritative_artifact": "...",
  "draft": "...",
  "critique": "...",
  "final_candidate": "..."
}

Installation

Use directly with npx (recommended, no install needed):

npx moltblock@0.11.0 "your task" --json

Or install globally:

npm install -g moltblock@0.11.0

Configuration

No configuration file is required. Moltblock auto-detects your LLM provider from environment variables and falls back to sensible defaults.

Optionally, place moltblock.json in your project root or ~/.moltblock/moltblock.json to customize model bindings:

{
  "agent": {
    "bindings": {
      "generator": { "backend": "google", "model": "gemini-2.0-flash" },
      "critic": { "backend": "google", "model": "gemini-2.0-flash" },
      "judge": { "backend": "google", "model": "gemini-2.0-flash" }
    }
  }
}

See the full configuration docs for policy rules and advanced options.

Source

• •Repository: github.com/moltblock/moltblock
• •npm: npmjs.com/package/moltblock
• •License: MIT

Security: Test Execution

When --test is used, moltblock writes LLM-generated code to a temporary file and runs vitest against it using the user-provided test file. This executes LLM-generated code in a Node.js process on the host machine. Mitigations:

• •The test file path must be provided explicitly by the user — moltblock does not select or generate test files
• •Generated code is written to os.tmpdir() and cleaned up after execution
• •Policy rules run before test execution to deny known dangerous patterns (e.g. rm -rf, eval, child_process, filesystem writes)
• •Without --test, no code execution occurs — only policy checks run against the generated artifact

Residual risk: Policy rules are pattern-based and cannot catch all dangerous code. LLM-generated code executed via --test may perform arbitrary actions within the permissions of the Node.js process. Users should review generated code or run moltblock in a sandboxed environment when verifying untrusted tasks.

Disclaimer

Moltblock reduces risk but does not eliminate it. Verification is best-effort — policy rules and LLM-based checks can miss dangerous patterns. Always review generated artifacts before executing them. The authors and contributors are not responsible for any damage, data loss, or security incidents resulting from the use of this tool. Use at your own risk.