15 GTM and Sales Hackathon Ideas You Can Build Without Writing Code

I have been reading through the GTM hackathons of 2026: the winners, the tracks, the starter kits, and the projects people posted afterwards. Two things stand out. The rooms are no longer full of engineers only. Sales reps, growth people, and MBA students now enter these events, and some organizers say plainly that you do not need to be technical. And almost every project, whoever builds it, stops at the same place: right before anyone talks to a customer.

That gap is the most interesting thing about GTM hackathons right now. The agents that hold the conversation, the ones that call the lead, run the demo, or rehearse the pitch, are rare at these events because a working voice agent used to need a developer stack. That is no longer true, and the projects that use this tend to have the best demos in the room.

This post is for the person who walks into a GTM or sales hackathon and wants to build something real in a day without writing code. I build Tough Tongue AI, where these agents run, so I will be specific about how each idea works. Most of the thinking applies whatever tool you pick.

What GTM hackathon teams built in 2026

The Agentic GTM Hackathon at Station F in Paris in July had about 60 teams building with Claude, Sillage, FullEnrich, and HubSpot. The winners were good projects: Gravity warms buyers up before outreach, Zorro turns buying signals into actions, and Net-Work writes a one-page brief before a sales call. Voice showed up only as an interface, for example a team approving actions by speaking.

The AI Growth Hackathon at YC ran for 24 hours in June with tracks like Sales Cyborgs, Revenue on Autopilot, and Reading Minds, aimed at GTM engineers who code. The Dublin edition of the Agentic GTM Hackathon in August gave teams Codex and Unify, told them they did not need to be technical, and suggested a lead scoring agent and a content engine built from customer calls. Berlin in October brought about 100 builders together around one idea: detect the signal, build the agent. A four-hour GTM Fellowship hack in San Francisco put its rule in its description: no slides, no fake demos, just working AI.

Outside public events, Ramp has talked about an internal hackathon project called Mo, a real-time copilot for reps on live calls, built by an account executive rather than an engineer. That one is closer to the conversation than most, and I think it is a sign of where these events are going.

Put the projects side by side and a pattern appears. Detect a signal, qualify it with a model, enrich the contact, write to the CRM, draft an email or a brief. It is a text pipeline, and it is useful. It also ends one step before the part of sales that is hardest to scale.

Why the conversation is still the open lane

Building an agent that talks has been a developer job. On a code-first framework, a voice agent starts with a repository, separate API keys for speech recognition, the language model, and the voice, a server to host it, a phone provider for calls, and a meeting-bot service if you want it in Google Meet. Each piece is well documented. Wiring all of them up, and making the agent handle interruptions without sounding broken, is more than most teams can finish between lunch and the demo.

So teams make a reasonable call. They build the parts they can finish, which are text, and stop before the voice.

What changed in the last year is that the voice layer can now be bought rather than built, and set up in plain language. You describe the agent: who it is, what it knows, what it should ask, when to hand off. It runs in the browser right away, and the same agent can take phone calls, join a Meet, or sit on a web page. If you do want to build the stack yourself, our comparison of LiveKit, Vapi, and Retell and the post on building versus buying a voice agent lay out what that takes.

The part that matters most for a non-engineer is where you do the building. Tough Tongue AI has an MCP server, so you can create and edit the agent from a chat in ChatGPT (there is an official ToughTongue AI app) or Claude (add it as a connector or install the plugin). You say “turn these three call notes into a skeptical CFO who pushes back on price” and the agent exists. The same chat can place a test call to your phone, send the agent into a Google Meet, and read back the scores. The MCP page has the setup for ChatGPT, Claude, Codex, Cursor, Copilot and Gemini, and the agents page shows what you can build.

15 hackathon ideas that hold the conversation

Each idea below ends in a moment a judge can feel in under a minute. That is the test I would use to pick one. If you cannot describe the stage moment in one sentence, the idea is probably still a pipeline.

They fall into three groups: agents that sell for you, agents that train your team, and agents that do GTM research. Most of them reuse the same building blocks, so once you have one working you can swap the channel or add a screen without starting over.

Agents that sell for you

1. Speed-to-lead caller. A new lead fills a form and gets a call within seconds. The agent says it is an AI assistant, names what the person asked about, asks three or four qualifying questions, and books a meeting or hands off. The form sends one request to the agent through a webhook, usually through a form tool or Zapier, and the call goes out from a phone number you connect. The research behind this is strong: the odds of reaching a web lead drop sharply after the first five minutes, which I covered in the speed to lead guide. The stage moment is the best one on this list: a judge types their own number into your form, and their phone rings before you finish your sentence.

2. AI demo rep in Google Meet. The agent joins a Google Meet or Zoom call as a participant, with an avatar face that shows inside the meeting. It greets the people in the call, presents a Google Slides deck and moves through it as they ask questions, and can switch to a live browser to show the product itself. Afterwards it books the next meeting or hands off to a human. For the demo, send the agent into a Meet the judges join from their own laptops. A face that greets them and then answers their questions about your product is hard to forget. The Google Meet and Zoom agent guide and the post on AI demo agents go deeper.

3. Closed-lost revival caller. Pull ten deals your team lost in the last year, with the loss reason for each. The agent calls the contact, mentions what has changed since (a new feature, a lower tier, a fixed bug), and asks whether the timing is better now. A typical hackathon pipeline would draft a re-engagement email at this point. The agent makes the call instead and records the answer. For the demo, a judge plays a lost buyer and you show the agent picking up the real loss reason from the CRM.

4. After-hours receptionist. Most small businesses miss calls after six. Give the agent an inbound number, your qualification rules, and a person to transfer to when a caller is ready to buy. For the demo, ask a judge to call the number “at midnight”. The agent answers, qualifies them, and transfers the call to a teammate standing at the back of the room.

5. Website concierge with a face. Embed the agent on a landing page with an avatar and let visitors ask about pricing by voice. Add a small live app beside it, such as a calculator that fills in as the visitor talks about their team size and current costs. This works well for a team that already has a product page and wants it to talk.

6. Same-day event follow-up. Conferences produce a spreadsheet of badge scans that nobody calls for a week. Load the sheet, have the agent call every lead before dinner with a short, specific opener (“you stopped by our booth about X”), and write the outcome back next to each row. If your hackathon is attached to a conference, use its own attendee list with permission, and the demo writes itself.

Agents that train your team

7. Objection gym from real calls. Paste three sets of call notes into ChatGPT or Claude and ask for the hardest buyer in them. The agent becomes that buyer, with the same objections in the same order, and a rubric scores every practice round. Some recipes for this chain four or five tools together by hand: a call recorder, a model, a voice service, and a scoring step. Here it is one chat. On stage, a teammate practices against the buyer live and the score appears when they hang up.

8. Persona mock ladder. Build one buyer per segment, each at three difficulty levels, from friendly to hostile, and score every rep on the same rubric so a manager sees who is ready. This is how one of our current customer conversations started: a sales leader at a large edtech company built parent personas for their admissions team at a hackathon this year, liked what they saw in a day, and is now looking at running them for the whole team. Hackathon builds that solve a manager’s real problem tend to outlive the hackathon.

9. Live deal-room board. Put a small app beside the conversation that the agent updates while it talks, for example a deal qualification board where each field fills in as the rep uncovers budget, the decision maker, and the pain, and the gaps stay red. On Tough Tongue AI these are called Agent Desktop apps: a small React app the agent can read and change during the session. The rep practices against an AI buyer, and the board shows what they missed before the call even ends.

10. Negotiation partner. An AI counterpart with a hidden walk-away point that it reveals only when the other side earns it. Salary talks, vendor renewals, and client pricing all work. A business school team built a version of this from scratch on a coding tool in a 48-hour hackathon. With the agent written in plain language, the same build fits in an afternoon, and phased disclosure (what the counterpart says only after the third good question) is a few sentences in the instructions.

11. Pitch rehearsal on your own deck. Give the agent your Google Slides deck and a buyer persona. It follows the deck as you present, and interrupts on the slide that invites the objection, the way a real buyer would on the pricing slide. The agent can also present the deck itself, which makes a good second half of the demo: first the human pitches, then the agent shows how it would have done it.

Agents that do GTM research

12. Voice customer discovery overnight. Write a discovery guide, load a list of people who agreed to talk, and let the agent run twenty short interviews while the team sleeps. Extraction pulls the same fields out of every call (current tool, biggest pain, what they would pay), so the morning starts with a sheet of answers instead of twenty recordings. The demo is the sheet, with one call played back to show the agent following up on a vague answer.

13. Win/loss interviewer. Lost deals rarely get a real debrief, because nobody wants to make that call. A five-minute call from an agent that asks “what tipped it” and then listens is easier for the buyer to accept, and the answers go straight into the CRM. Judges who have run sales teams will recognize the problem immediately.

14. Message tester with buyer personas. Write three value propositions and five buyer personas. Each persona hears each pitch, reacts in character, and a rubric scores how well the message landed. The output is a scoreboard of which message wins with which buyer, which is a real input for positioning work and a clean thing to show on a slide.

15. Brief, then rehearse. One of the Paris winners wrote a one-page brief before a sales call. Take the next step: the brief becomes a buyer persona, and the rep gets a five-minute rehearsal with that exact buyer before the real call. Through MCP, the assistant that wrote the brief can create the persona in the same chat.

The building blocks, so you can mix your own

The fifteen ideas are combinations of a small set of blocks. Knowing them lets you change an idea to fit the track you are in, or the data your sponsors hand out.

A few practical notes on each layer. Phone calls go out from a number you own, connected through a SIP trunk from a provider such as Twilio, Plivo, or Telnyx. Set that up the night before, because it is the one step that needs an account with another company. Meeting agents join Google Meet and Zoom. Mid-call actions use custom functions, which are plain HTTP calls to any API, so anything with an endpoint (HubSpot, your own backend, a spreadsheet script) is reachable. After the call, every conversation is transcribed and scored against a rubric you write, and the fields you ask for come back as structured data you can send anywhere with a webhook.

If your track gives you data from a sponsor, such as enrichment or CRM records, the cleanest pattern is to let the existing pipeline do its job and hand its output to the agent as variables. The pipeline finds the right person and writes the context. The agent has the conversation.

What judges remember

Judges at these events see dozens of demos in an afternoon, and the rule many organizers state openly is that they want working software, not slides. A pipeline demo usually ends in a dashboard: enriched rows, a score, a drafted email. It may be the better piece of engineering, and it still asks the judge to imagine the outcome.

The demos that land put the judge inside the product. Their phone rings, or they join a Meet and something greets them. They do not have to imagine anything, because they just had the conversation.

Live demos fail more often than slides, so plan for it:

• Record a clean run of the full demo the evening before, and keep it ready to play.
• Bring a second phone, and test the venue network with a real call an hour before judging.
• Write a short script for the judge (“say you are the VP of sales at a 50-person company”) so the conversation goes where your demo shines.
• Keep the agent’s opening line under ten seconds. Long intros are where live demos lose the room.

A good test the night before is to hand your phone to someone who has not seen the project and watch their face during the first twenty seconds of the call. If they lean in, keep it. If they look at you, rewrite the opening.

Getting started, and hackathon credits

The fastest path on the day looks like this. Connect Tough Tongue AI to ChatGPT or Claude from the agents page. Describe the agent in the chat and talk to it in the browser until the first minute sounds right. Then connect the channel your demo needs: a phone number, a Meet link, or an embed on your page. Spend whatever time is left on the opening line, the stage script, and the backup recording.

Every account starts with 25 free minutes in the browser, which is enough to build and test the agent. Phone calls and meeting agents are on paid plans, starting with Starter at $20 a month, and students get 50% off Starter with a .edu email (pricing). We do not run a standard hackathon credit program, but if you are building at a hackathon, tell us about it through the hackathon credits form. I read every request and reply by email.

The short version

GTM hackathons in 2026 are full of good pipelines that find the right buyer and write the right message. Very few of them talk to the buyer. That last step used to need a developer stack, and now it needs a clear description of the conversation and an afternoon of testing.

If you are walking into a GTM or sales hackathon this season, pick the idea whose stage moment you can describe in one sentence, build it from a chat, and spend the rest of the day making the first twenty seconds of the conversation feel right. The judges will remember the call.