TaScan Local Agent

Run Claude Code on your machine. Trigger tasks from anywhere.

Drop a task into the TaScan AI Inbox — from your phone, the MCP server, the API, or the cloud agent — and this local agent picks it up, spawns Claude Code, does real filesystem work, and marks it complete. You get an email. Task. Scan. Done.

Part of the TaScan Protocol — the universal task infrastructure platform built by Love Technologies. 10 patents filed, 265 claims.

What It Does

The TaScan Local Agent is the on-machine half of Patent 2: Closed-Loop Autonomous Operations Protocol. It bridges the gap between cloud-triggered tasks and local execution — the thing that actually needs your filesystem, your codebase, your installed tools.

The loop:

1. Agent subscribes to Supabase Realtime on the AI Inbox task list
2. A new task arrives (from mobile, API, MCP, cloud agent — anywhere)
3. Agent spawns claude -p with the task as the prompt
4. Claude Code reads files, edits code, runs commands, deploys — real work
5. Output is captured and saved as a task completion record in Supabase
6. Admin receives email notification via SendGrid
7. Every execution is logged to a daily tracking list (data flywheel)

A fallback polling loop (every 30 seconds by default) catches anything Realtime missed.

Task Routing

Tasks are routed by prefix in the title:

| Prefix | Handled By | |--------|-----------| | CODE: | Local agent (Claude Code) | | SHELL: | Local agent (Claude Code) | | PLAN: | Local agent (Claude Code) | | (no prefix) | Local agent (Claude Code) | | RESEARCH: | Cloud agent (agent-execute.js) | | WRITE: | Cloud agent (agent-execute.js) | | REVIEW: | Cloud agent (agent-execute.js) | | MCP: | Cloud agent (agent-execute.js) |

Both agents watch the same inbox. The local agent claims CODE/SHELL/PLAN tasks first if it is running. Cloud handles everything else instantly.

Safety

Blocked keywords (task is rejected, not executed):

• drop table
• rm -rf
• format c:
• truncate
• shutdown
• reboot
• private key

Dry run mode executes zero real commands — use it to verify wiring before going live.

Setup

Prerequisites

• Node.js >= 18
• Claude Code CLI (claude) installed and authenticated
• A TaScan account with an AI Inbox task list

Install

npm install -g tascan-local-agent

Or run from a local clone:

git clone https://github.com/mikelove2nd/tascan-agent
cd tascan-agent
npm install

Environment Variables

| Variable | Required | Description | |----------|----------|-------------| | SUPABASE_SERVICE_KEY | Yes | Supabase service role key (not the anon key) | | SUPABASE_URL | No | Defaults to the TaScan Supabase instance | | AI_INBOX_ID | No | UUID of your AI Inbox task list | | CLAUDE_WORKER_ID | No | UUID of the Claude worker record in Supabase | | WORKING_DIR | No | Directory for Claude Code to work in (defaults to parent of agent dir) | | DRY_RUN | No | Set to true to log tasks without executing | | CLAUDE_TIMEOUT | No | Max ms to wait for Claude Code (default: 300000 = 5 min) | | POLL_INTERVAL | No | Fallback poll interval in ms (default: 30000 = 30 sec) |

Run

export SUPABASE_SERVICE_KEY="eyJ..."
npm start

Dry run (no real execution):

npm run dry

Or if installed globally:

SUPABASE_SERVICE_KEY="eyJ..." tascan-agent

How It Integrates

The local agent is one piece of the TaScan closed-loop architecture:

Mobile / API / MCP / Claude Desktop
         |
         v
   TaScan AI Inbox (Supabase)
         |
    +----|----+
    |         |
    v         v
Local Agent  Cloud Agent
(CODE/SHELL) (RESEARCH/WRITE)
    |         |
    v         v
 Claude    Claude API
  Code     (Haiku/Sonnet)
    |         |
    +----+----+
         |
         v
  task_completions (Supabase)
         |
         v
  Email notification (SendGrid)

The MCP server (tascan-mcp) can dispatch tasks to the AI Inbox directly. Any MCP-connected client — Claude Desktop, Cursor, your own tool — can trigger local code execution on your machine with a single tool call.

Patent 2: Closed-Loop Autonomous Operations Protocol

This package is a reference implementation of Patent 2. The core claim: a system where tasks originating from any channel (human, API, AI agent) are routed to the appropriate execution environment, executed autonomously, and fed back into a structured completion and audit record — creating a closed loop between intent and action.

The data flywheel: every execution creates a log entry. Every log entry trains the routing logic. The system gets better the more it runs.

Part of the TaScan Ecosystem

• TaScan App: https://app.tascan.io
• MCP Server: npm install -g tascan-mcp (32 tools, OAuth 2.0, remote endpoint)
• The Council (LLM Arena): https://llm-arena-love.netlify.app — 9 AI models deliberating in real time
• Patents: 10 provisional patents filed, 265 claims (Patents 1-10, filed 2026)

Author

Michael Edward Love II Founder, Love Technologies / TaScan Protocol [email protected] | https://tascan.io

License

MIT