Inspiration Most AIs can only use whatever tools their developers shipped, and asking for something new gets you "I can't do that." I wanted an assistant that could truly extend itself — one that, when I asked for a skill it lacked, would design, write, test, and install the tool itself, live, behind an approval gate. What it does Ember is a self-improving AI assistant. You ask for a capability it doesn't have → it plans the tool → writes the Python code → tests it in an isolated sandbox → and installs it as a brand-new live skill you can use in chat (with your approval at every step). I built several this way: Hackathon Buddy (countdown + feasibility), Idea Research (live GitHub + Hacker News search), Browser Buddy (open apps/websites/search), and an interactive Pomodoro Timer (custom HTML/JS UI). Clay-gamified chat (XP, levels, avatar), voice input, and a multi-provider LLM router (Gemini, OpenRouter, OpenCode) under the hood. How we built it Python FastAPI chat server + React/TS/Vite frontend + a dedicated tool-runtime service that runs every forged skill in its own .venv. The forge loop: plan → code → sandbox-test → contract-test → verify → install (no-cache — the file is the API). Built during STEMist Hacks IV, with two tools forged live on camera. Challenges we ran into Reliability: free Gemini quota (20/day) + truncated model outputs that split the tool code from its tests. Schema bug: a subtle malformed required flag made the provider reject every valid tool with a cryptic 400 — fixed with a recursive sanitizer. Keyless research: DuckDuckGo rate-limited us and Bing returned junk, so I pivoted to the GitHub Search API and HN Algolia with a points>30 filter. Live-demo risk: every step had to be resumable because a single failure killed the recording. Accomplishments that we're proud of Watched Ember delete a tool and rebuild it end-to-end on camera; shipped four working skills each with unit tests verified in the sandbox before install; made keyless web research actually return real results; got Browser Buddy to discover and launch installed apps via Windows registry/Start-Menu. What we learned Build recovery into the loop, not just the first run — when quota and truncation hit, a clean retry path is what turns a one-shot trick into a product. Sandbox everything before trusting it. And "honest" > "optimistic": Hackathon Buddy tells you when tasks won't fit instead of padding the schedule. What's next for Ember Interactive tool templates (calendar/list/custom apps), more research sources (arXiv, Reddit), persistence for tool outputs, batch forging of multiple tools, container/Docker deployment, and multi-user auth for safe internet exposure.

Built With

Share this project:

Updates

Submission history