TRUEFOUNDRY × POLARIS HACKATHON

Build Agents That Act. Not ones that just answer.

Anything can talk. An agent has to reach your real systems, run the code it writes without breaking anything, and know when to stop and ask. One day, one room in Bangalore, built on TrueForge — TrueFoundry's open-source agent harness. Round 1 is open to everyone online; shortlisted builders get a seat at the Polaris campus on 26 September.

SAT 26 SEP 2026 ₹3,00,000 PRIZE POOL BANGALORE · POLARIS CAMPUS IN-PERSON · ONE DAY TWO ROUNDS · FREE
Register for Round 1 ↗ See what you'll build
event.status
// WHAT IT IS

One day, one room, 350–400 builders

TrueFoundry and Polaris are co-hosting a day-long hackathon in Bangalore, built on TrueForge, the open-source agent harness TrueFoundry released in August.

If you followed the online Agent Harness Hackathon earlier this year, this is the same challenge in a very different format: seven hours instead of seven days, in one room instead of on your own, with mentors from the team that built the harness sitting at the next table. Bring the idea you did not have time to finish, or a new one.

Entry runs in two rounds. Round 1 is an open online form: tell us why you want in and what you plan to build. We shortlist from those answers. Round 2 is the build itself, in person, on 26 September, using TrueForge and TrueFoundry's AI gateway alongside whatever else your stack needs.

No prior TrueFoundry experience is required — TrueForge is MIT-licensed and runs locally with one command. Bring an idea and a laptop.

DATE
Sat, 26 Sept 2026
VENUE
Polaris campus, Bangalore
FORMAT
In-person, one day
SCALE
350–400 builders
ENTRY
Free · shortlisted
PRIZE POOL
₹3,00,000
// THEME

Agents That Act

Reach real systems. Run real code. Stop before it hurts.

A model that answers questions is a feature. A model that opens the pull request, runs the migration, or revokes the access key is something else entirely — and the distance between the two is not a better prompt. It is plumbing.

Acting on the world takes three things a chat window never needed, and every one of them is infrastructure rather than intelligence. That plumbing is the theme. Pick a job in any domain that is genuinely worth handing to an agent, and build the thing that does it end to end.

REQUIREMENT 01
It has to reach something real
Your database, your repo, your cloud account, your ticket queue. Connected over MCP, with real credentials and real consequences — not a mocked function returning fixture data.
REQUIREMENT 02
It has to run what it writes
An agent that reasons about code but cannot execute it is guessing. Generated code needs somewhere to actually run — isolated, disposable, and unable to damage anything when it is wrong.
REQUIREMENT 03
It has to know when to stop
Some actions cannot be taken back: the email sends, the row drops, the key is revoked. The agent should pause at that line and wait for a person, every time, and make it obvious what it is about to do.

The layer that handles all three sits between the model and everything it touches, and building it yourself would eat the whole day. TrueForge is an open-source one: MCP tools, sandboxed execution, approval checkpoints, subagents, and sessions that survive a reconnect. You write the agent; the harness does the machinery. TrueFoundry's AI gateway is there too if you want budgets, rate limits, and traces without touching your agent code — useful, but not required to win anything.

// WHAT YOU'LL BUILD

Six agents worth building

Starting points, not categories. Any domain is fair game — what matters is that the agent reaches something real and has a line it will not cross alone. Each of these has one.

01 · EASIEST START
Cloud cost janitor
Hunts down idle instances, orphaned volumes, and forgotten load balancers, works out what they cost you a month, and drafts a teardown plan you can read before anything disappears.
REACHES: your cloud billing + infra APIs GATE: deleting a resource
02 · DATA
Migration rehearsal agent
Takes a schema change, restores a copy into a sandbox, runs the migration there, diffs what actually happened to the rows, and reports back before it goes anywhere near the real database.
REACHES: your database GATE: applying to production
03 · SHIPPING
Release captain
Reads every commit since the last tag, runs the test suite in a sandbox, writes release notes a human would actually publish, then holds at the one step you cannot undo.
REACHES: GitHub + a package registry GATE: tagging and publishing
04 · SUPPORT
Ticket resolver
Picks up a bug report, tries to reproduce it in a sandbox, and comes back with either a patch and a draft reply or an honest "could not reproduce, here is what I tried".
REACHES: Linear, Jira, or Zendesk GATE: replying to the customer
05 · SECURITY
Access reviewer
Walks your IAM roles and service accounts, finds the permissions nobody has used in ninety days, and proposes a least-privilege diff with the blast radius of each revocation spelled out.
REACHES: your identity provider GATE: revoking access
06 · HERO PROJECT
Runbook executor
Reads a runbook written for a human and actually executes it, step by step, doing the reversible parts itself and stopping at every step that changes the world. The hardest of these, and the most useful.
REACHES: your infrastructure GATE: every destructive step

What every submission needs

Five things. The first one decides whether a project qualifies at all — if it would work just as well as a text box, it is not this hackathon.

  • An agent running on TrueForge, with the harness visibly doing the work: a real tool reached, code executed in the sandbox, a pause before something irreversible.
  • One job, finished. Seven hours is short. One narrow task done end to end beats three half-built features every time.
  • The approval moment in your demo. Show us where the code ran and show us the agent stopping to ask. This is the part everyone forgets to film.
  • A public repo with a README that works on someone else's laptop. Note which AI assistants you used while building.
  • Only what is yours to connect. Your accounts, your data, your keys — and none of those keys in the repo or the demo video.
// THE STACK

Start before the day

You do not have to wait for 26 September to touch the tooling. TrueForge is open source, MIT-licensed, and runs on your laptop in one command.

TrueForge is the runtime layer that turns a model into a working agent: 40+ built-in tools, MCP servers with OAuth, sandboxed code execution, tool-approval checkpoints, subagents, git-backed skills, and persistent sessions. Local mode is a single process on SQLite — no infrastructure to stand up.

# Node.js 22+ · runs locally, no infra $ npx @truefoundry/trueforge

Point it at any model provider — OpenAI, Anthropic, Gemini, or any OpenAI-compatible endpoint — or route everything through TrueFoundry's AI gateway to get budgets, rate limits, and traces without touching your agent code. Walking in on the day already familiar with the harness is the single biggest advantage you can give yourself.

GITHUB REPO ↗ QUICKSTART DOCS ↗ WHAT IS TRUEFORGE ↗ ASK IN DISCORD ↗
// HOW IT WORKS

Two rounds

Round 1 is online and open to all. Round 2 is in person, by invitation.

01
Round 1 — register online
A short form: who you are, why you want to join, and what you plan to build. No prototype needed at this stage. Takes about five minutes. Register here ↗
02
Post about your build — ₹75,000 rides on this
Share what you are building on LinkedIn or X and tag @truefoundry and @polariscodes. Two community prizes of ₹50,000 and ₹25,000 go to the best public build stories — the write-up, thread, or clip that best explains the job you gave the agent, how you wired it up, and what surprised you. Optional, but it is a quarter of the prize pool.
03
Shortlist
We read every entry and shortlist on the strength and specificity of the build idea, not credentials. Seats are capped at 350–400. Shortlisted builders get an email invitation with venue details.
04
Round 2 — build in person
Saturday 26 September at the Polaris campus in Bangalore. Build on the day, demo to judges, ship something that runs. Food, power, and wifi are on us.
05
Demo and judging
Live demos at the end of the day. Working code beats slides — expect to show the agent reaching a real system and stopping to ask before it does the irreversible thing. See the full rubric ↓
// DAY-OF AGENDA

Saturday, 26 September

Provisional running order, all times IST. Expect these to shift by up to thirty minutes; the final schedule goes out with the shortlist invitation.

09:00
Check-in and setupBadges, wifi, workspace. Get your environment running before the clock starts.
10:00
Kickoff and problem briefThe three requirements in detail, the rubric, and what counts as the harness doing real work.
10:30
Technical walkthrough: TrueForge + AI gatewayConnecting MCP tools, running generated code in the sandbox, and wiring an approval gate. Gateway credentials handed out here.
12:00
Build starts · lunch servedMentors from TrueFoundry and Polaris on the floor for the rest of the day.
16:00
Mentor checkpointOptional review of your approach and failure-mode coverage.
19:00
Submissions closeCode and a short write-up on the hackathon platform.
19:30
Demos and judgingShortlisted teams demo live to the judging panel.
21:00
Results and prizesWinners announced on stage.
// PRIZES

₹3,00,000, two ways to win

Cash, paid out after the event. Build prizes are judged on the day. Community prizes are judged on what you post publicly in the run-up.

₹3,00,000
TOTAL PRIZE POOL · ALL CASH · INR
// BUILD PRIZES · ₹2,25,000
FIRST PLACE
₹1,00,000
Winning team, judged on the day
SECOND PLACE
₹75,000
Runner-up team
THIRD PLACE
₹50,000
Third-placed team
// COMMUNITY PRIZES · ₹75,000
BEST BUILD STORY
₹50,000
Best public write-up, thread, or demo video about the job you handed over and how you wired it up
RUNNER-UP BUILD STORY
₹25,000
Second-best public build story

Community prizes are open to every Round 1 registrant, shortlisted or not — post on LinkedIn or X, tag @truefoundry and @polariscodes, and you are in the running. Build prizes are for teams at the venue on 26 September. Partner credits and category awards may be added; nothing already listed here will shrink.

// JUDGING

How you'll be scored

Published up front so you can build against it. Every submission is scored on all five, out of 100.

30
The harness is doing the workThis is the one that decides whether you qualify. A judge has to watch TrueForge reach a real tool, run generated code in the sandbox, and hold for a person. An agent that is really a prompt with a nice wrapper scores near zero here regardless of how good the rest is.
25
It actually runsWorking software, not a deck. Someone who has never seen the project should be able to clone it, follow the README, and get it going on their own laptop. Narrow scope that works scores above broad scope that doesn't.
20
Where it stopsWhich actions did you decide the agent may never take alone, and can you defend the line you drew? We look at what is sandboxed, what is gated, how clearly the agent explains what it is about to do, and how small the damage would be if it got something wrong.
15
A job worth handing overWould a real person actually delegate this, and is it an interesting thing to delegate? Solving a chore someone genuinely has beats an impressive demo of nothing in particular.
10
Demo clarityFive minutes to show the job, the agent doing it, and where the harness fits. Judges will ask you to explain your own architecture, so be able to.

The judging panel is engineers and developer-relations leads working on agents in production, drawn from TrueFoundry, Polaris, and the wider ecosystem. Names go up here as invitations are confirmed.

// PARTNERS

Partner with us

TrueFoundry and Polaris are funding this between them. We are talking to model and compute providers about credits for participants, and logos go up here only once they are signed.

SPONSORSHIP · OPEN
Support 350–400 builders in Bangalore
Model credits, compute, hardware, or a track prize. Reach out to the hosts on LinkedIn.
// WHO SHOULD ATTEND

Eligibility

A GOOD FIT IF
  • You have shipped, or are shipping, something with an LLM in the loop.
  • You write code day to day — backend, ML, platform, or full-stack.
  • You have a chore at work you have quietly wanted to hand to an agent.
  • You are a student or new grad who can build without hand-holding.
  • You can be in Bangalore, in person, on Saturday 26 September.
THE RULES
  • Free to enter and attend. No fee at any stage.
  • Open to anyone 18+ who can attend in person. No student-only restriction.
  • Round 1 registration is individual. Teams of up to 4 form on the day or in advance in Discord.
  • Build on the day. Pre-built projects are not eligible; prior research and reading the docs are encouraged.
  • Your agent must run on TrueForge. Everything else — language, framework, model provider — is your call.
  • AI assistants are allowed and expected. Disclose what you used in your README, and be ready to explain your own architecture. Submissions nobody on the team can explain are not eligible.
  • A short code of conduct applies: be decent to each other, to mentors, and to the venue. Harassment means you leave.
  • You keep full IP in what you build. We ask only to showcase your demo.
// KEY DATES

Timeline

OPEN NOW
Round 1 registration is liveRoughly five minutes. Review happens on a rolling basis, so earlier is better.
FRI 18 SEP 2026
Round 1 closes23:59 IST. No extensions — seats are allocated well before this if we fill up.
MON 15 SEP 2026
Shortlist announcedInvitations by email, with venue address and logistics. Check spam.
SAT 26 SEP 2026
Round 2 — hackathon dayPolaris campus, Bangalore. Build, demo, results the same evening.
// FAQ

Questions

Does it cost anything?
No. Round 1 registration and the in-person day are both free. Food and wifi are covered on the day. Travel to Bangalore is on you.
Can I take part remotely?
Round 1 is fully online. Round 2 is not — it is in person at the Polaris campus in Bangalore on 26 September. If you cannot attend that day, please do not take a shortlist seat.
How big can a team be?
Up to four people. Round 1 registration is individual — everyone applies for themselves — and teams form either in our Discord beforehand or in the room on the day. Solo entries are welcome and have won hackathons before.
Do I have to use TrueForge?
Yes — your agent has to run on TrueForge, the open-source harness this event is built around. Beyond that nothing is prescribed: any language, any framework, any model provider. TrueForge is MIT-licensed and runs locally with npx @truefoundry/trueforge, so you can get familiar with it today rather than learning it under time pressure.
Can I use Claude, Copilot, or Cursor to write the code?
Yes, and most of you will. We only ask two things: note what you used in your README, and be able to explain your own architecture when a judge asks. A project nobody on the team can walk through is not eligible, regardless of how well it runs.
How do the community prizes work?
₹50,000 and ₹25,000 go to the two best public build stories — a LinkedIn post, an X thread, a blog post, or a demo video that explains what you built, how you wired it up, and what surprised you. Post it, tag @truefoundry and @polariscodes, and you are entered. Open to every Round 1 registrant whether or not you get a shortlist seat, which means you can win from anywhere.
Who can apply?
Anyone 18 or over who writes code and can attend in Bangalore on the day. Students, working engineers, and founders are all welcome. No TrueFoundry experience required.
How does shortlisting work?
We read your answers on why you want to join and what you plan to build. A specific, technically plausible idea beats a polished résumé. Posting publicly about your build helps. Seats are capped at 350–400.
What should I bring?
A laptop and charger, a phone hotspot as backup, your own API keys if you have preferred providers, and a government photo ID for campus entry. Power and wifi are provided.
Who owns what I build?
You do. You keep full IP in your project. We ask only for permission to reference and showcase your demo when we write about the event.
Where do I register and submit?
Registration is on our event page. Project submission happens on the day — a repo link and a short write-up, details in the kickoff brief. Everything else runs through Discord.
// HOSTS

Who is running this

TrueFoundry
TRUEFORGE · AI GATEWAY
TrueFoundry builds infrastructure for running AI in production: an AI gateway in front of models for routing, rate limits, cost control, and observability, and TrueForge, the open-source agent harness that runs and instruments agents. TrueForge went public in August — this hackathon is what comes next.
Polaris
CO-HOST · VENUE
Polaris is co-hosting and providing the campus in Bangalore for the build day, along with mentors on the floor. Find them at @polariscodes.

Find a team in Discord

Round 1 is an individual application, so the Discord is where you find the other three people. It is also where mentors answer TrueForge questions before the day, where the schedule gets posted, and where we announce the shortlist.

Join the Discord ↗
// ROUND 1

Register for Round 1.

Tell us what you want to build. A paragraph is enough. We review on a rolling basis and email shortlist invitations directly.

Round 1 closes Friday 18 September. Seats for the in-person day are capped at 350–400 and allocated by review, not by who clicked first — but a specific idea submitted early gets read when we are still fresh.

Read the judging rubric before you write your answer. It tells you what we are looking for.

What the form asks

  • Who you are, and where you write code.
  • What you plan to build — the job you want to hand over, and what it has to reach.
  • Why you want the seat.
  • Whether you already have a team, or want one.
Register for Round 1 ↗
FREE · INDIVIDUAL APPLICATION · CLOSES 18 SEP 2026
TEAMS OF UP TO 4 FORM LATER