Everything

Inspiration We kept seeing people struggle with the same problem: they'd describe what they wanted automated — "when X happens, do Y and notify me" — but actually building it meant stitching together APIs, handling auth, writing glue code, deploying it somewhere. We thought: what if you could just say what you want, and an AI agent builds the whole automation for you? Not just the logic, but the actual runnable workflow with real tools, schedules, and delivery.

What it does ImagineLabs turns plain-English into multi-step automations. You describe what you want — "search Reddit for posts about my competitor, summarize the sentiment, text me the findings" — and it generates a workflow with sources, AI reasoning steps, tool actions (via Composio), and delivery. You can trigger automations manually, schedule them (cron), or run them from a Telegram bot. There's also a Brain AI that you can just talk to — it can list your automations, run them, or build new ones on the fly. We even built hooks so an external Guild.ai agent can orchestrate ImagineLabs, creating an "agent that builds agents."

How we built it TypeScript monorepo (Next.js 14 web app, worker engine, scheduler, Telegram bot). The generator uses Claude to turn prompts into workflow graphs, the engine executes nodes (RSS, web scrape, LLM steps, HTTP, Composio tools), and Postgres stores everything. We integrated Composio's SDK for OAuth + 1000+ app tools, wired Pioneer for multi-model routing (cheap tasks → OSS models, reasoning → Claude), and built a full orchestration API (list/create/run) so external agents can drive it. The Telegram bot long-polls and bridges to the Brain, so you can run automations from your phone.

Challenges we ran into OAuth state everywhere. Reddit worked first try, but figuring out why the UI showed "not connected" even after successful OAuth took hours — turned out to be a webpack serialization quirk in how Composio's SDK sent array query params.

SSRF. Since the generator is LLM-driven and the hooks API is internet-exposed, we had to add egress guards on all the fetch nodes (http/web.scrape/rss) to block private/loopback/link-local IPs. One adversarial review finding turned into a real security fix.

Telegram fail-open. The bot initially answered anyone if you set a token without a chat ID. We caught it in review and made it refuse to start instead — authentication bugs are scary.

Accomplishments that we're proud of The meta layer. We didn't just build an automation platform — we built a platform other agents can drive. The Guild integration is an external agent that lists, builds, and runs automations via API. An agent orchestrating an agent factory felt like the future.

Real tools, real runs. This isn't a demo with hardcoded responses — it actually calls Reddit's API via Composio OAuth, runs Claude for reasoning, scrapes real web pages, and texts real results to Telegram. Every run is multi-step and deterministic.

The Telegram UX. Being able to text "run the market intelligence briefing" from your phone and get a real AI-synthesized report back is wild. It made the whole thing feel like a product, not just a hackathon project.

What we learned Composio's SDK is powerful but young. The toolkit catalog is huge and managed auth works great, but the documentation gaps (how userIds actually work, what headers the bundled SDK sends) meant we had to probe the API directly and read OpenAPI schemas.

LLM-generated workflows need guardrails. The generator occasionally picks the wrong Reddit tool (newest posts instead of search), and we saw it almost let an SSRF through. You can't just trust the LLM output — you need schema validation + runtime guards.

Multi-agent orchestration is real. The Brain can create and run agents, the Telegram bot drives the Brain, and Guild can orchestrate the whole thing. We accidentally built three layers of agency, and it actually worked.

What's next for ImagineLabs More Composio apps. We have Reddit working; Gmail/Slack/Notion/Sheets are one OAuth click away. A five-app workflow (fetch emails → filter with AI → save to Notion → post in Slack → add to calendar) would be a killer demo.

Pioneer live. The routing is wired (cheap tasks → OSS models), but we need the billing plan active. Once it's on, extraction steps run 30× cheaper while reasoning stays on Claude.

Langfuse observability. Trace every LLM call (generation, Brain reasoning, every workflow node) so you can see the full agent decision tree.

Self-healing workflows. If a step fails, have the Brain analyze the error and rewrite the node config to fix it — turning runtime failures into learning moments.

Built With

• composio
• guildai
• pioneer
• senso