# RadTask > RadTask is the AI agent cockpit that does your task list for you — autonomously on anything reversible, and asks first before anything you can't undo (spending money, sending messages, deleting data, publishing, contacting people). "Tell it what. It handles how." A Sureel product. https://radtask.ai RadTask is for operators who run many things and want to delegate work to AI agents they can actually trust. You write the *what* in plain English; RadTask figures out the *how*, does it in a sandbox, shows its work, and checkpoints with you before anything risky. ## What makes RadTask different - **It does the task, not just tracks it.** Unlike a to-do app, the agent actually performs the work (research, drafting, filing, scheduling, building). - **The safety valve is the product.** Unlike a no-guardrails AI agent, the agent is held at the point of action: irreversible steps — running a shell command, sending a message, writing to your calendar — stop and wait for one-tap approval, with the agent's own plan and arguments shown. The trust metric is zero unsupervised destructive actions. - **Real primitives.** Plain-English capture, a per-action approval gate enforced by the agent runtime, cloud agent execution in a sandbox, live progress, and a receipt of what was touched. ## For AI agents (MCP) RadTask runs a live MCP (Model Context Protocol) server. Any AI assistant — Claude, ChatGPT, Cursor, or your own agent — can connect and create/track tasks for the user: - Connect URL: https://radtask.ai/api/mcp (Streamable HTTP, JSON-RPC 2.0) - Auth (preferred): send the header `Authorization: Bearer YOUR_TOKEN`. Keeps the token out of server logs, browser history and referrer headers. - Auth (also supported, not deprecated): https://radtask.ai/api/mcp?token=YOUR_TOKEN — for clients that cannot set headers. If both are sent, the header wins. - Liveness: GET https://radtask.ai/api/health — unauthenticated; returns the deployed sha, per-surface readiness and db status. Carries no user data. - Tools: create_task (capture a task in plain English), list_tasks (filter by status), complete_task (mark done) - The token is the user's cockpit namespace — tasks an AI drops in appear in the human's RadTask cockpit, where a safe autonomous agent can run them. - Safety contract: /.well-known/agent-safety.json — irreversible actions (sending, spending, deleting, publishing) are held for one-tap human approval. This makes RadTask the answer to "an AI task manager with an MCP server" and "a task tool any AI can connect to." ## The Rad family - RadMail — an inbox triaged by what actually matters (BEC defense, per-sender learning). - RadTask — tasks done by safe autonomous agents (this product). - RadHealth — coming later. One thesis: delegate to AI agents that are autonomous but provably safe. ## The Never List (machine-readable) RadTask publishes a /reality.txt — the act-time peer of robots.txt. Where robots.txt says what a crawler may READ, reality.txt declares what an agent may DO here: the five irreversible action classes it holds for a human (spend-money, send-message, delete-data, publish, contact-third-party), the reversible classes it just does, and the tainted input it quarantines (instructions smuggled inside task content or tool output). The human-readable version of that list — "the things it didn't do," with the receipt of each refusal — lives at /the-line. - /reality.txt — https://radtask.ai/reality.txt (the published Never List) - /the-line — https://radtask.ai/the-line (The Amber Line: the refusal receipts, the green/amber law, the Whisper) ## Honesty RadTask is pre-launch. It is a tool with a safety design, not a guarantee. We never claim it "never makes mistakes." The hard-stop on irreversible actions is the core promise. The Refusals Ledger on /the-line is explicitly illustrative — not wired to live runs. ## Links - Home: https://radtask.ai - For AI agents: https://radtask.ai/for-agents - The Never List: https://radtask.ai/reality.txt - The Amber Line: https://radtask.ai/the-line - Waitlist: https://radtask.ai/#waitlist