2026-01-16 · Authensor

How to Add AI Agent Safety to Sublime Text

SafeClaw by Authensor provides deny-by-default AI agent safety for Sublime Text users. If you use AI-powered plugins like Sublime Copilot or custom LLM integrations, SafeClaw gates every action through your policy before execution. It logs all decisions to a hash-chained audit trail, supports both Claude and OpenAI, and is backed by 446 tests.

Prerequisites

Step 1: Install SafeClaw

Open Sublime Text's built-in console (Ctrl+ ) or your system terminal and run:

npx @authensor/safeclaw

This creates the .safeclaw/ directory in your project root with a default deny-all policy and empty audit log.

Step 2: Define Your Policy

Create .safeclaw/policy.yaml:

version: 1
default: deny


rules:
  - action: file.read
    paths:
      - "src/**"
      - "lib/**"
      - "tests/**"
    decision: allow



- action: file.write
    paths:
      - "src/**"
      - "lib/**"
    decision: prompt



- action: shell.execute
    decision: deny



- action: network.request
    domains:
      - "api.openai.com"
      - "api.anthropic.com"
    decision: allow

Step 3: Create a Sublime Build System

Sublime Text build systems let you run SafeClaw commands with a keyboard shortcut. Create a file at ~/.config/sublime-text/Packages/User/SafeClaw.sublime-build (Linux) or ~/Library/Application Support/Sublime Text/Packages/User/SafeClaw.sublime-build (macOS):

{
  "shell_cmd": "npx @authensor/safeclaw audit --tail 10",
  "working_dir": "$project_path",
  "selector": "",
  "variants": [
    {
      "name": "Verify Chain",
      "shell_cmd": "npx @authensor/safeclaw audit --verify"
    },
    {
      "name": "Status",
      "shell_cmd": "npx @authensor/safeclaw status"
    },
    {
      "name": "Validate Policy",
      "shell_cmd": "npx @authensor/safeclaw policy --validate"
    }
  ]
}

Access variants with Ctrl+Shift+B (or Cmd+Shift+B on macOS) and select the SafeClaw command you need.

Step 4: Add a Sublime Project Configuration

In your .sublime-project file, add SafeClaw settings:

{
  "folders": [
    {
      "path": "."
    }
  ],
  "settings": {
    "safeclaw.enabled": true,
    "safeclaw.policy_path": ".safeclaw/policy.yaml"
  },
  "build_systems": [
    {
      "name": "SafeClaw Audit",
      "shell_cmd": "npx @authensor/safeclaw audit --tail 10",
      "working_dir": "$project_path"
    }
  ]
}

Step 5: Wrap AI Plugins with SafeClaw

If your AI agent plugin runs as a subprocess, wrap it:

npx @authensor/safeclaw wrap -- python3 ai_plugin_server.py

For Sublime LSP-based AI integrations, configure the server command in your LSP settings to run through SafeClaw:

{
  "clients": {
    "my-ai-agent": {
      "command": ["npx", "@authensor/safeclaw", "wrap", "--", "node", "ai-agent-lsp.js"],
      "selector": "source.python | source.js"
    }
  }
}

Step 6: Test the Integration

Trigger an AI agent action that attempts to write a file. SafeClaw should intercept it, apply the prompt decision, and log the result. Check:

npx @authensor/safeclaw audit --tail 5

Each entry shows the action type, file path, decision, timestamp, and hash chain link.

Summary

SafeClaw adds a safety layer to Sublime Text through build systems, project configuration, and plugin wrapping. The deny-by-default model ensures AI agents cannot act outside your policy. Hash-chained audit logs provide tamper-evident evidence of every action. SafeClaw is MIT licensed and open source.

Related Guides

Try SafeClaw

Action-level gating for AI agents. Set it up in your browser in 60 seconds.

$ npx @authensor/safeclaw