I built a desktop AI agent that writes its own tools — behind verification gates

Short summary

AMA-teras is an open-source desktop AI agent (Electron + TypeScript, AGPL) that writes its own tool plugins when it encounters capability gaps, with every generated tool passing typecheck, unit tests, smoke runs, and human approval before promotion. The agent supports swappable models (Anthropic, OpenAI, Moonshot) and enforces protected paths to prevent self-modification of its approval mechanism. The author shares failure stories including an incident where the agent deleted its own work branch, leading to auto-backup safeguards.

• •AMA-teras is an open-source desktop AI agent that generates its own tool plugins under strict verification gates: typecheck, tests, smoke run, human approval
• •Supports swappable LLM providers (Anthropic, OpenAI, Moonshot) and enforces protected paths preventing self-modification of the approval engine
• •Includes nightly autonomous mode on separate branches, phone-based approval, and community plugin verification-evidence records

Generated with AI, which can make mistakes.