BuildBridge

The Problem

You bookmark 10 tutorials a week. You build maybe one a year. Not because you're lazy — because the gap between reading a great article and creating a repo, setting up the structure, and typing the first command is just large enough to kill momentum every single time.

The article is right there. The steps are all listed. But getting from "that looks useful" to "I have a working folder open" takes 15–20 minutes of copy-paste, file creation, and hunting for that one dependency version. By then you've lost the energy that made you want to do it.

BuildBridge closes that gap. Paste the URL. In under 30 seconds, you have a GitHub repo with the full implementation kit — every step, every command, every prerequisite — ready to clone and start. The article stays open. The momentum doesn't die.

What It Does

Three ways to trigger it

How It Works

Architecture

Three things can trigger the pipeline: the web app, the Firefox extension, or an Ambiguous @mention (via a Next.js webhook route). YouTube URLs take a detour through Whisper transcription before hitting the main pipeline. Everything else goes straight to Analyze Content.

The distribution layer fires all six outputs simultaneously — not sequentially. GitHub repo creation, Slack notification, and the three Ambiguous actions (channel message, task, document) all go out in parallel using Promise.all. Trigger.dev schedules the follow-up separately via its background job API.

Key files

The CopilotKit Fix

The AI sidebar was supposed to be the easiest part. CopilotKit provides a prebuilt sidebar component — you register tools with useCopilotAction, the LLM calls them, and you're done. Except in v1.71.1, that's not how it works anymore. And nothing in the docs tells you that.

The root cause: CopilotKit v1.71.1 uses the ag-ui v2 protocol internally, which tracks tool calls by ID and expects a TOOL_RESULT event to be emitted back into the stream. The old useCopilotAction hook runs the handler but never emits that event. So the runtime keeps waiting. The LLM retries. No response ever arrives.

// ❌ Old (v1) — handler runs but result never reaches ag-ui stream
useCopilotAction({
  name: "list_ambiguous_channels",
  parameters: [{ name: "channelId", type: "string" }],  // array format
  handler: async () => { ... }
});

// ✅ New (v2) — result emitted through ag-ui TOOL_RESULT event
import { useFrontendTool, useDefaultRenderTool } from "@copilotkit/react-core/v2";
import { z } from "zod";

useDefaultRenderTool();  // required — handles rendering for all tools

useFrontendTool({
  name: "list_ambiguous_channels",
  parameters: z.object({ channelId: z.string() }),  // Zod schema required
  handler: async ({ channelId }) => { ... }
}, []);

Three things change: the import path (@copilotkit/react-core/v2 not the main package), the parameters format (Zod schema instead of a plain array), and the addition of useDefaultRenderTool() once in the component. After that, the sidebar works — tool calls, results, text responses, all of it.

Live tool call — "List all my Ambiguous channels" → real workspace data, no blank screen.

Build Story

The hackathon was announced with a 24-hour window. The theme: Agents, Everywhere — Bots, Channels & More. Sponsors: Ambiguous, Nebius AI, CopilotKit, Exa, Trigger.dev. The challenge was to build something that lived inside a workspace environment, not just notified it.

The core pipeline — URL to GitHub repo — came together in the first few hours. Exa's API is straightforward; the LLM prompt required three iterations to get reliable step extraction. GitHub repo creation via Octokit was four lines. Slack took ten minutes. Ambiguous was where things got interesting.

The Ambiguous integration had three parallel goals: post a channel message, create a task, create a searchable document. All three fire simultaneously on every "Build Kit" click. That part worked cleanly. What didn't work was the @mention webhook — Ambiguous's cloud servers can't POST to localhost, which requires either ngrok or the Ambiguous CLI dev tunnel. The webhook code is complete but the demo runs via the web UI and direct API calls instead.

The CopilotKit sidebar was the last thing built and the last thing fixed. The "Tool result is missing" error appeared 2 hours before the deadline. The fix — migrating from useCopilotAction to useFrontendTool — took 45 minutes to identify and 10 minutes to apply once the cause was clear.

The demo video was assembled with ffmpeg and Puppeteer in 20 minutes. No audio — a walkthrough script was written instead. The submission went in with 15 minutes to spare.

Key Learnings

Moments In Between

Platform Review

Ambiguous

CopilotKit

Nebius AI (LLM) + OpenRouter (fallback)

Exa

Team

Attempted to recruit on UX and backend before the hackathon. Didn't come together in time. The full stack — frontend, pipeline, integrations, demo video, submission — was built solo.