HackStack

Project Info

Crewly

Devpost

Inspiration

Tradespeople lose hours every week to “back office” work: calls/texts, finding leads, checking whether a job is even feasible, and tracking supplies. We wanted to build something that makes a solo contractor feel like they have an office manager so they can spend their time on the work they’re actually paid (and proud) to do.

What it does

Crewly is an AI-powered operations layer for blue-collar pros. Our framework: Poke/Message onboard tradesperson orchestrates everything Voice Agent handles incoming calls from potential clients schedules appointments gives quotes based on materials and situation Inventory Agent keeps track of current inventory with cost looks at cheapest prices using BrightData reorders low inventory using BrowserBase Stagehand Outreach Agent looks on sites like Nextdoor and Craigslist to get leads on potential clients using BrowserBase Stagehand Marketing Agent using OpenAI API to generate business cards and brochures Logging Agent keeps track of client correspondence and customer satisfaction Mapping Agent feasibility and ordering of tasks tracking scope of outreach Website Agent uses Vercel and OpenAI API to spawn and deploy company websites Document Agent generates and signs documents

How we built it

We built Crewly as a tool-driven agent system: a chat interface orchestrates an MCP tool server that exposes business actions like feasibility checks, pricing floor calculation, and inventory logic. We used structured tool inputs/outputs so the agent can reliably chain steps: collect job details → validate feasibility → estimate a minimum viable price → propose next actions. We also added Modal configuration within our MCP to allow for our agents to deploy secure and concurrent sandboxes to run code in.

Challenges we ran into

Tool wiring and reliability: making sure the agent passes the right units/types (miles vs meters, hours vs minutes). Environment + auth issues: handling missing API keys and avoiding circular imports in a multi-module tools package. Especially in handling environment with fragile dependencies, it was easy to run into issues when one part of the workflow was not properly set up. Keeping outputs safe and consistent: ensuring tools return predictable outputs.

Accomplishments we're proud of

A working end-to-end flow from a single message to actionable ops steps. Modular tools that are easy to extend (distance validation, concurrent sandbox runs, inventory hooks). Clean tool semantics + documentation so the agent behaves more like a real assistant and less like a demo.

What we learned

Tool design matters as much as the model: clear names, tight schemas, and good docstrings dramatically improve correctness. “Ops” is a chain, not a single feature—small automations compound into real time saved. Debuggability is everything: structured outputs and health checks make iteration fast.

What's next

Generalize beyond trades into any contractor's staff team to lead intake, quoting, scheduling, follow-ups, lightweight CRM, and parts/procurement plus tighter integrations (maps, calendars, messaging) so the AI can run more of the workflow autonomously.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • AnthropicIn code
  • CSSIn code
  • FastAPIIn code
  • HTMLIn code
  • JavaScriptIn code
  • PythonIn code
  • TypeScriptIn code
  • OpenAIClaimed
  • VercelClaimed

7 of 9 appear in the indexed code. 2 claimed on Devpost could not be matched to code, which may simply mean the tool leaves no trace in the repository.

AI coding agents

No AI coding agent signals were found in this repository.

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

152 KB

Source files

26

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars