Hack for Humanity London. Build the planning tools your community needs.
Pick a real problem from your community. Plan it with a ready-made Timefold model, or build your own with an AI agent, then ship an app people can use. One day, teams of two to five.
Corgi Cafe
74 Great Eastern St, London EC2A 3JL
Thursday, October 29, 2026
09:00 to 15:00 GMT
With The AI Collective
Hosted by The AI Collective with Timefold.
Six hours to get to a working solution
Briefing
Problem briefing and a walkthrough of both build options. Teams form.
Hack
Five hours to build your model and your app.
Lunch is served
Hacking continues.
Demos and judging
Five minutes per team, live. Judges score, the room votes.
Close
Winners announced.
Build an app people can use, with Timefold as the optimization engine
Somewhere near you, a hospital is struggling to plan its staff rotas fairly. A food bank is working out who drives where. A volunteer group is juggling too many people and too few time slots.
These are all planning problems. Pick one from your own community and build an app that solves it, with Timefold doing the planning. A website, a mobile app, an AI agent workflow: the form is up to you.
The app
Whatever people in your community use: a website, a mobile app, an AI agent workflow that calls the model, a chatbot. Any language or framework works. A single HTML page is enough.
The planning model
The engine behind the app. It takes your data and returns the best plan it can find. Use a ready-made model or build your own.
Realistic data
Made-up but believable people, places, time slots, and rules from your community, at the size of the real problem.
Your data goes in, the best plan comes out
You describe the problem
What gets planned, what each thing can be assigned to, and the rules, sent as JSON. Hard constraints never break. Soft constraints are preferences.
Timefold searches
It tries many possible plans, scores each one against your constraints, and keeps the best. The longer it runs, the better the plan gets. It runs in the background, and you can stop it whenever the plan is good enough.
You get the best plan back
The assignments as JSON, with a score that says how good the plan is, such as 0hard/-12soft. 0hard means every hard constraint holds, and closer to 0soft is better. Your app shows the plan to people.
POST /rotas
{
"volunteers": [
{ "name": "Amira", "skills": ["driving"], "maxHours": 6 },
{ "name": "Tom", "skills": ["packing"], "maxHours": 4 }
],
"shifts": [
{ "id": "sat-09", "start": "09:00", "end": "12:00", "skill": "driving" },
{ "id": "sat-13", "start": "13:00", "end": "16:00", "skill": "packing" }
]
}200 OK
{
"assignments": [
{ "shift": "sat-09", "volunteer": "Amira" },
{ "shift": "sat-13", "volunteer": "Tom" }
],
"score": "0hard/0soft"
}At the end of the day, you demo the app live: solve a realistic dataset, change something, and show the plan adapt.
Community planning problems, with the constraints that make them interesting.
The best projects start from a problem someone on the team has seen up close. If you need a starting point, these are all planning problems a morning is enough for.
Fair rotas for a care team or hospital ward
What gets planned: shifts, each assigned to a staff member.
Food-rescue and grocery deliveries
What gets planned: pick-ups and drop-offs, each assigned to a driver and a position in their route.
Practice schedules for a local sports club
What gets planned: practice sessions, each assigned to a pitch and a time slot.
Volunteer sign-ups for a community event
What gets planned: volunteer tasks, each assigned to a volunteer.
Lifts to medical appointments
What gets planned: patient journeys, each assigned to a volunteer driver.
Room bookings for a community center
What gets planned: activities, each assigned to a room and a time slot.
Each of these maps to a type of problem Timefold already handles at scale, so your agent has plenty to work with. The suggested starting points are just that: if your version of the problem has rules a ready-made model can't express, build your own.
Teams of two to five, formed on the day.
Arriving alone is normal: most people do.
How teams form
What every project needs to qualify.
Every project needs to follow these rules to be judged. If something isn't covered, ask an organizer before you build it.
Help people locally
Your project solves a real planning problem in a local community, with the purpose of benefiting people.
Use Timefold for the planning
Your solution includes a Timefold planning model that produces the plan: a ready-made model on the Timefold Platform, or your own.
AI agents are welcome
Use Claude Code, Codex, or any other agent as much as you like. Your team is responsible for everything you build, so understand what your code does.
Open tools and data are fine
Starter projects, open-source libraries, and public datasets are allowed. Credit them.
No real personal data
Don't use names, addresses, health details, or other personal data of real people. Use made-up or anonymized data.
Follow the code of conduct
All attendees agree to The AI Collective's code of conduct and privacy policy.
Disqualification
The organizers can disqualify a project that breaks these rules, for example one that uses real personal data. Teams get a chance to explain before any decision.
Use a ready-made model, or build your own.
Both give you a Timefold planning model behind a REST API. Pick one as a team.
Use a ready-made model
Call a model Timefold already runs in production. Fit your data to it, and spend the day on your app.
You need a free trial account. No Java.
Build your own model
Describe your problem in plain language. PRD Chat turns it into requirements, and a quickstart gives you working code to adapt.
You need a JDK 21 or higher, Git, and an AI agent.
Custom models can score highest on modeling the problem. Not sure which to pick? Each problem idea suggests an option, and mentors help at the briefing.
Get a model, then build the app.
Follow the steps for the option your team picked. Both end with a model you can call from your app.
Call a ready-made model
The Timefold Platform hosts ready-made planning models behind a REST API. The model's rules are already written and tested. Your job is to describe your problem in the shape the model expects.
Prerequisites and API key
What gets planned
Fits problems like
Field Service Routing (FSR)
Visits, each assigned to a person or vehicle and a place in their route. Handles time windows, skills, travel time, and working hours.
Food-rescue rounds, home care visits
Employee Shift Scheduling (ESS)
Shifts, each assigned to an employee. Handles skills, availability, rest time, contracts, and fairness.
Care rotas, volunteer shifts at an event or a food bank
Pick-up and Delivery Routing (PDR)
Jobs that pick something up in one place and drop it off in another, each assigned to a vehicle. Handles time windows and vehicle capacity.
Lifts to medical appointments, moving donated goods between sites
Task Scheduling
Tasks, each assigned to a person and a start time. Handles skills, durations, and availability.
Repair jobs for a community workshop, prep tasks before an event
Each model has its own guide in the Timefold documentation at docs.timefold.ai, with the full request format and examples. The Platform also lists other models you can try. For questions about a model's input, output, or API, point your agent at docs.timefold.ai/llms.txt, which indexes the whole documentation for AI agents.
Know what the model can and can't do. A ready-made model supports a fixed set of rules, which you switch on and tune through the input. If a rule your problem needs isn't supported, say so in your demo, or switch to option B. Ask a mentor if you're unsure.
When the model returns a plan you trust, check it with the questions under option B, step 2, then build your app around it.
Prerequisites and the MCP server
You need a Java Development Kit (JDK), version 21 or higher, and an AI coding agent that supports MCP servers, such as Claude Code or Codex, and Git. Install the JDK with one method for your operating system.
Command
macOS, with Homebrew
brew install --cask temurin@21
Windows, with winget
winget install EclipseAdoptium.Temurin.21.JDK
Ubuntu or Debian
sudo apt install openjdk-21-jdk
Any OS, with SDKMAN
sdk list java to find a 21.x (or newer) Temurin identifier, then sdk install java <identifier>
If java -version shows an older version, your terminal is picking up a different JDK. Ask a mentor, or ask your AI agent to fix your JAVA_HOME.
You don't need Maven or Gradle separately. Most Java projects, including the Timefold quickstarts, ship a wrapper script (./mvnw or ./gradlew) that downloads the build tool for you.
Connect the MCP server
The Timefold MCP server, PRD Chat, gives your AI agent the Timefold knowledge and tools it needs to turn your description into requirements and a planning model. You only need to connect it once.
Get the server URL. Open prd.timefold.ai and copy the MCP server URL shown there. Use it wherever <MCP_SERVER_URL> appears below.
How to add the server
Claude Code
Run the Claude Code command from a terminal, inside your team's project folder. Start Claude Code with claude, then type /mcp. You should see timefold-prd listed as connected. If the server asks you to authenticate, follow the prompt from the /mcp screen.
Codex
Add the Codex config to ~/.codex/config.toml, then restart Codex. If your Codex version uses a different configuration format, check the Codex documentation or ask a mentor.
Other agents
Any agent that supports remote MCP servers over HTTP works. Add a server called timefold-prd with the URL from prd.timefold.ai.
Check that it works with the prompt in the Ask your agent block. If the agent lists the Timefold tools, you're ready to build your model.
What to try
The server doesn't show up
Check you ran the command in the right folder, then restart the agent.
The server shows as failed
Check the URL for typos and that the venue Wi-Fi is connected.
The agent ignores the tools
Mention the server by name in your prompt: "Use the timefold-prd MCP server to..."
Anything else
Raise a hand. A Timefold mentor will come to your table.
brew install --cask temurin@21winget install EclipseAdoptium.Temurin.21.JDKsudo apt install openjdk-21-jdksdk list java # find a 21.x (or newer) Temurin identifier
sdk install java <identifier>java -version
# openjdk version "21.0.x" or higherclaude mcp add --transport http timefold-prd <MCP_SERVER_URL>[mcp_servers.timefold-prd]
url = "<MCP_SERVER_URL>"Which tools does the timefold-prd MCP server give you?
Summarize what each one is for.Build your model from a quickstart
A planning model knows what gets planned, what the options are, and which rules decide whether a plan is good. You describe those three things. Your agent and the MCP server turn them into a working Timefold model with a REST API, starting from a quickstart.
The Timefold quickstarts are complete, working Timefold projects, each with a REST API and a simple web page. Pick the one closest to your problem and use it as your team's base.
Start from
People assigned to shifts
Employee scheduling
Drivers visiting stops in order
Vehicle routing
Sessions assigned to a room or pitch and a time slot
School timetabling or meeting scheduling
Command
Clone the quickstarts
git clone https://github.com/TimefoldAI/timefold-quickstarts.git
Open the quickstart you picked
cd timefold-quickstarts/java/employee-scheduling
Run it
./mvnw quarkus:dev
Open http://localhost:8080 and solve the example data once, so you know the base works. Then copy the quickstart into your team's own repository and adapt it from there. Each quickstart's README has the exact steps.
Answer these questions first
Before you write a prompt, agree on the answers as a team.
What gets planned?
The things you assign: shifts, deliveries, practice slots.
What are the options?
What each thing can be assigned to: a volunteer, a driver and a time, a pitch and a time.
What are the rules?
Split them into hard and soft constraints.
Meaning
Example
Hard constraint
Must never break. A plan that breaks one is infeasible.
A driver can't be in two places at once.
Soft constraint
A preference. The model tries to break as few as possible.
Spread weekend shifts fairly between volunteers.
Start with three to five constraints. A small model that works beats a big one that doesn't. You can add more once the first version solves.
Questions about Timefold Solver? Use llms.txt. When your agent gets stuck on how to write a constraint or configure the solver, point it at docs.timefold.ai/llms.txt, which indexes the whole Timefold documentation for AI agents.
Test it like the people who'll use it
Once the model returns a plan, check it with whoever on the team knows the problem best.
Ask the model to explain itself. Ask your agent to show the score broken down by constraint. It's the fastest way to see why the model made a choice, and it makes a good demo moment.
Does the plan break any hard constraint?
If the score shows a negative hard number, something is infeasible. Ask the agent which constraint breaks and why.
Would a real person accept this plan?
If the answer is "no, because...", that "because" is a missing constraint.
Does it scale to a realistic size?
Try a dataset the size of the real problem, not just five items.
Common pitfalls
Rules that are really data
"Amira can only work Saturdays" is availability data on Amira, not a separate constraint per person.
Everything as hard constraints
If nothing can be bent, the model often finds no feasible plan. Make preferences soft.
Personal data
Use made-up names and addresses, as the rules require.
When the model returns a plan you trust, build your app around it. Next: read how the demo works.
git clone https://github.com/TimefoldAI/timefold-quickstarts.git
cd timefold-quickstarts/java/employee-scheduling # or the quickstart you picked
./mvnw quarkus:devThere's no written submission. Your five-minute live demo is your entry.
The judges and the audience score what they see. Hacking stops when demos begin. Each team gets five minutes. The order is announced after lunch. When your time is up, the next team starts, so practice once with a timer.
What goes on your slides
Time
What to cover
The problem
1 minute
Who has it, where, and what it costs them today
The rules
1 minute
The three or four constraints that matter most
The live demo
2.5 minutes
Solve a realistic dataset, then change something and solve again
What's next
30 seconds
Who would use this on Monday, and what's missing
Show, don't tell. Judges score the working solution live. A small app that solves in front of the room beats a long explanation of a bigger one.
Half the score is the panel. Half is the room.
Each project's final score is 50% judges panel and 50% audience vote. Both parts are converted to 100 points, so they count equally.
Judges panel
Four weighted criteria, scored one to five.
Audience vote
Each team ranks its top three. Not your own team.
Judges panel (50%)
Each judge scores every project on four criteria, from one to five.
Weight
What judges look for
Community impact
30%
A real, local problem. Clear who benefits and how much it helps them.
Modeling the problem
30%
The plan follows the real-world rules of the problem. Hard and soft constraints are chosen sensibly, and the plan is feasible and makes sense. Custom constraints written for this community's specific rules score highest.
Working solution
20%
The model and the app work end to end, live, on realistic data.
Demo
20%
The problem, the solution, and the result are clear within five minutes.
Custom models can score highest on modeling. A well-configured ready-made model (option A) can score up to a three or four on modeling the problem. A five needs custom constraints written for your community's specific rules (option B).
Scores a one
Scores a three
Scores a five
Community impact
Generic problem, no clear user
A real problem, but the users and the benefit stay vague
A named kind of group with a problem the team understands first-hand
Modeling the problem
Toy rules, or the plan breaks hard constraints
A ready-made model, well configured, with the data mapped sensibly and a feasible plan
Custom constraints written for this community's specific rules, a feasible plan, and the team can explain a trade-off the model made
Working solution
Screenshots or mock-ups only
Works end to end, but on toy data or with manual steps
Live app calling the model, reacting to a change in the data
Demo
Hard to follow, runs over five minutes
Clear, but the result or the "why" gets lost
Problem, model, and result in five minutes, with one memorable moment
Audience vote (50%)
Each team votes once, as a team. After the last demo, agree on the three projects you'd most like to see used in your own community, rank them, and hand your top three to an organizer. You can't rank your own team. A first choice earns 3 ranking points, a second choice 2, and a third choice 1.
Ties
If two projects have the same final score, the one with more judges points wins. If they're still tied, the higher community impact score decides.
€500 in Amazon vouchers
The winning teams may also be invited to showcase their projects in The AI Collective's global newsletter, which reaches more than 300,000 subscribers.
Where to get help on the day, and answers to common questions.
Mentors
There's no online support channel. Timefold engineers walk around the room during the hacking time, so raise a hand or wave one down.
Documentation questions
Ask your agent, and point it at docs.timefold.ai/llms.txt.
Wi-Fi
Provided by Corgi Cafe. Network details are shared at the venue.
Yes. If you know a local problem inside out, you're as valuable to a team as the person writing code. You decide which rules matter and check whether the plan makes sense to real people.
Your team does. Keep building it after the event, open-source it, or hand it to the community group it was built for.
Start with the Timefold documentation and the quickstarts. For your AI agent, use docs.timefold.ai/llms.txt.
If your problem is mostly routing or shift scheduling, option A gets you a strong plan fastest, and leaves more time for the app. If your problem has its own kind of rules, or doesn't fit either model, pick option B. Ask a mentor at the briefing if you're unsure.
No. Option A runs on a free trial account. Sign up at app.timefold.ai.
Hack for Humanity London
Thursday, October 29, 2026, 09:00 to 15:00 GMT, at Corgi Cafe, 74 Great Eastern St, London.