Plain Ballot

Plain Ballot

Every race on your ballot, matched to what you care about. Every position backed by a quote you can check.

plainballot.com · How it works · The open data · Contribute

Who's running, and where do they stand? Pick the issues you care about, set a dial for each, and enter your address. Plain Ballot goes through your ballot and shows who fits you, why, and where that came from. Party labels stay hidden until you ask. No ads, no account, and your address isn’t stored.

Covering every race in California, plus U.S. House and Senate races in all 50 states, for Nov 3, 2026.

Two projects, both free and open:

• [Plain Ballot](https://plainballot.com) (this repo): the voter guide.
• [Open Election Data](https://github.com/n8peace/open-election-data): the research behind it. Every candidate on the same 18 issues, each position with an exact quote and its source, researched by AI agents that have to agree. Free to use, including commercially, under the Open Database License, with a free API.

AI that has to show its work

Chatbots get voting facts wrong. Plain Ballot is built so no single model's word ever reaches a voter.

• Agents have to agree. Every candidate is researched by independent agents on different models (Claude Code, Codex, GPT-5.6 Luna), each searching and reading on its own. A position is published only when they agree: 2 of 3 with no conflict, or a clear majority of 10 when they don't.
• Every quote is verified. Each claim carries an exact quote, fetched from its source and matched word for word, or it's thrown out. A nightly job rechecks every published quote.
• No guessing. An agent not finding a source never counts as a vote. No majority means the issue stays blank, and the ballot says so.
• Matching isn't AI. It's plain, deterministic arithmetic in your browser (lib/match.ts). Same answers, same ballot, every time.
• Voters can push back. "Not right?" on any claim files a public issue and reruns the research with fresh agents. Corrections are public.
• Every change is checked for bias. Three models from different companies review each pull request for loaded wording and one-sided logic, and tests require that matching treats both ends of every dial the same.
• Every word is public. The dial wording, the agreement rules and every researched position live in Open Election Data; the matching lives here.

Use the data

The research is its own project: [Open Election Data](https://github.com/n8peace/open-election-data). It has a free, static API with no key needed:

curl -s https://n8peace.github.io/open-election-data/v1/divisions/ca/cd-10.json

Use it for anything, including commercial products, under the Open Database License: credit it, and share improvements to the data under the same terms.

Send your agent

Plain Ballot works without a browser. Point any AI agent that supports MCP (Claude, ChatGPT, Cursor and others) at:

https://plainballot.com/mcp

It has three read-only tools:

• list_issues: the 18 issues and how each side is worded.
• get_ballot: every researched race for a street address, with each choice's position, the exact quote, and the source.
• score_fit: how closely each choice fits the priorities the voter gave, using the same arithmetic as the site. No model is involved.

The agent is told the same rules the site follows: stay nonpartisan, quote and link sources, never fill in a missing position, and only score priorities the voter actually stated. Addresses aren't stored.

Prefer plain JSON? GET https://plainballot.com/api/ballot?address=... returns the same ballot, and the research itself is a free static API.

What already exists, and why we still built this

Good sources are already out there, and Plain Ballot learns from them:

• [Ballotpedia](https://ballotpedia.org) is the most complete record of who's running, and the research agents read it. Its bulk data is a paid product, and its pages aren't ours to republish.
• [BallotReady](https://organizations.ballotready.org/ballotready-api) offers candidate issue stances through a paid API for organizations.
• [Vote Smart](https://www.votesmart.org/votesmart-api) collects positions through its Political Courage Test, a survey candidates fill in, along with voting records and ratings.
• [Google Civic](https://developers.google.com/civic-information) and [Democracy Works](https://democracy.works/elections-api) cover who's on the ballot and how to vote.

What we couldn't find was a free, open dataset of where candidates stand, with a checkable quote behind every position, with every candidate on the same issues so they can be compared, and free for anyone to reuse. So we built one: Open Election Data. Every claim still comes from the candidate's own words: agents follow a summary back to the candidate's site, voting record or interview and quote that.

Research your state

The research runs on your own Claude or ChatGPT plan, so anyone can add coverage without paying for API credits. It happens in Open Election Data, which has an open "research your state" issue for each state.

Run it locally

npm install

cp .env.example .env.local npm run dev # http://localhost:3000

Without keys, it runs on a fictional sample ballot and the dials still work. Researched races come from the open data's API (set ELECTION_DATA_DIR to a local checkout's data/positions to work offline). Keys turn on typed priorities (AI_GATEWAY_API_KEY), real ballot lookup (GOOGLE_CIVIC_API_KEY) and address suggestions (GOOGLE_PLACES_API_KEY).

npm test               # matching, neutrality, sharing, abuse limits

npm run eval # does the AI put people on the right side of each dial? (calls the model) npm run check:issues # dial wording matches the research data

How it's built

Next.js on Vercel. The Census geocoder finds a voter's districts. Researched races come from the Open Election Data API, matched to voters by district. BotID, rate limits and a spending cap protect the paid endpoints (docs/security.md). The full picture, with a diagram: docs/ARCHITECTURE.md. Coding with an AI agent? Point it at AGENTS.md.

Contributing

The most useful help is research and checking research, and neither needs code. See CONTRIBUTING.md, or open a ready-to-run copy in your browser:

License

AGPL-3.0. Anyone can run their own copy, but a hosted copy has to publish its changes, so nobody can quietly run a biased version.

Star history