The same code runs a small team's day-to-day tracker and the public demo (with DEMO_MODE=1).
Because your progress should be exponential, not linear.
- Live demo: https://exponential-demo.onrender.com
- Live site of author: https://wheatlay.is-a.dev
A Linear-style issue tracker built around Claude Code. Issues, sub-issues, priorities, labels, sprints with a review, meetings and polls, a realtime board. The difference is the loop around the AI agent:
- Issue. Someone files a bug or an idea.
- Agent analysis. An agent reads the code and writes a structured analysis (sections plus decisions) back to the issue.
- Human decision. People click an option, or write their own answer. The agent never decides alone.
- Agent ships. It starts from a ready prompt, commits are linked to the issue by its ID, and deploys are tracked.
- Change notes. A short HUMAN note and a detailed technical note for the next agent. Together they write the sprint review.
No sign-up: the landing page starts a guided 5-step tour on issue ACME-12 (Analyze, read the analysis, pick a decision, Start work, follow the activity and the Sprint review).
- The demo is a shared sandbox and resets every hour.
- The AI agent in the demo is a replay of a recorded session. No model is called.
- The service runs on Render's free plan and sleeps when idle, so the first request can take up to a minute.
Three optional layers (see docs/agent.md for what each installs and how to remove it; release notes in CHANGELOG.md):
- MCP server at
/mcpover Streamable HTTP, signed in with OAuth in the browser (API tokens still work). Tools list, read and save issues, comments, relations, analyses and change notes; the analyze and close procedures are served asget_procedureand MCP prompts. - Claude Code plugin served by the tracker itself (
plugin/): thin skills that fetch the procedure from the server, plus a worklog hook that reports files and commits (never the conversation). Every machine gets the same version from the server. - Structured analysis via the
set_analysistool: sections and decisions with options and a recommendation, rendered in the UI as buttons. - Start work puts a ready prompt (editable per project) into Claude Code.
- Commits are linked to issues by ID (
ACME-12) from a GitHub push webhook, only for the repository configured in the project. - Deploy events come from a generic webhook, GitHub Deployments, or optionally Coolify.
- exponential-agent (
agent/, single exe) runs analysis on the user's machine with Claude Code or Codex, per-computer repo mappings and a tool allow-list, so malicious issue text (prompt injection) cannot run arbitrary commands or read secret files.
- Fastify + TypeScript, run directly by
tsx. - Postgres + Drizzle ORM;
LISTEN/NOTIFYis turned into Server-Sent Events for realtime updates. - React 19 + TanStack Query + Vite; drag and drop with dnd-kit.
- Web Push notifications, iCal feed, Zod validation.
- Single instance by design: timers and presence live in the process.
src/server/service.ts- issue, status, label and activity logic.src/server/rules.ts- domain rules (Done requires change notes, identifier parsing, commit references).src/server/mcp.ts- the MCP server and its tools.src/server/analysis.ts- structured analysis and decisions.src/server/relations.ts- blocks, duplicates and related links.src/server/sprints.ts- sprint lifecycle and review.src/server/deploys.ts- deploy status tracking.src/server/hooks.ts- GitHub and deploy webhooks, commit linking.src/server/demo/- demo mode: seed data, simulated agent, hourly reset.src/web/Board.tsx- kanban board with drag and drop.src/web/IssuePage.tsx- the issue view with analysis and decisions.plugin/- Claude Code plugin: skills and worklog hook.agent/- exponential-agent (standalone exe).tests/- Vitest suites against a real Postgres.
docker compose -p tracker up -d # Postgres on localhost:5434
npm ci
npm run dev # API on :3000, migrations run at start
npm run dev:web # UI on :5173, proxied to the APILog in without GitHub: put DEV_LOGIN=1 in .env (npm run dev reads .env; copy it from .env.example) and just open http://localhost:5173: you are signed in automatically as the first active admin (or first active user). /auth/dev?login=you signs in as someone else, /login shows the normal login screen. Ignored when NODE_ENV=production.
Sign-in methods: GitHub, Google, and e-mail + password. New people get in through a one-time project invite link (Settings › People › Invite link; valid 7 days) or are added by an admin by GitHub login or e-mail. A workspace admin can create a one-time password link for an account (Settings › Accounts › Password link).
Demo mode: start the API with DEMO_MODE=1. Sample data for an empty database: npx tsx scripts/seed-demo.ts.
Copy .env.example. Only DATABASE_URL is needed to start; everything else enables a feature.
| Variable | Required | Description |
|---|---|---|
DATABASE_URL |
yes | Postgres connection string (default postgres://tracker:tracker@localhost:5434/tracker). |
BASE_URL |
for login, MCP | Public URL; used for the OAuth callback and links in MCP/plugin output. |
PORT |
no | HTTP port (default 3000). |
NODE_ENV |
no | production disables DEV_LOGIN. |
DEV_LOGIN |
no | 1 enables /auth/dev?login=... for local development. |
GITHUB_CLIENT_ID, GITHUB_CLIENT_SECRET |
for login | GitHub OAuth App; callback ${BASE_URL}/auth/callback. |
ADMIN_LOGINS |
no | Comma-separated GitHub logins that become admins on first login. |
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET |
for Google login | Google OAuth client (OIDC, scopes openid email profile); redirect URI ${BASE_URL}/auth/google/callback. Without GOOGLE_CLIENT_ID the Google routes and button are off. Only verified e-mails log in. |
ADMIN_EMAILS |
no | Comma-separated Google e-mails that become admins on first login. |
APP_TIMEZONE |
no | IANA time zone for meeting hours and default sprint review (default UTC). |
APP_LOCALE |
no | Language of server-sent messages (Discord, push, reminders): en (default) or pl. |
DEMO_MODE |
no | 1 runs the public demo: guest sessions, tour, simulated agent, hourly reset. |
DEMO_RESET_MINUTES |
no | Demo reset interval in minutes (default 60). |
PUBLIC_REPO_URL |
no | Repository link shown on the landing page. |
GITHUB_WEBHOOK_SECRET |
for commit linking | HMAC secret of the GitHub webhook (/hooks/github, push and deployment_status). |
GITHUB_API_TOKEN |
for private repos | Server-wide token for GitHub API calls (deploy commit ranges, commit authors, repo check). A project can set its own token in Settings › Integrations › GitHub repo, which wins over this one. Without any token a private repo deploy marks only the deployed commit, and Settings › Deploys shows a warning. |
DEPLOY_WEBHOOK_SECRET |
for deploy webhook | Secret sent in X-Webhook-Secret to /hooks/deploy. |
COOLIFY_API_URL, COOLIFY_API_TOKEN, COOLIFY_WEBHOOK_SECRET |
optional | Coolify deploy source. |
VAPID_PUBLIC_KEY, VAPID_PRIVATE_KEY, VAPID_SUBJECT |
for Web Push | VAPID keys and contact (mailto:) for push notifications. |
UPLOADS_DIR |
no | Where uploaded images are stored (default data/uploads). |
LOG_LEVEL |
no | Fastify log level (default info). |
A project with deploy source "webhook" accepts deploy events from any CI:
curl -X POST "$BASE_URL/hooks/deploy" \
-H "X-Webhook-Secret: $DEPLOY_WEBHOOK_SECRET" \
-H "Content-Type: application/json" \
-d '{"project":"ACME","sha":"'"$GITHUB_SHA"'","status":"finished"}'status is one of queued, in_progress, finished, failed. As a GitHub Actions step:
- name: Report deploy to Exponential
if: always()
run: |
curl -fsS -X POST "${{ secrets.TRACKER_URL }}/hooks/deploy" \
-H "X-Webhook-Secret: ${{ secrets.DEPLOY_WEBHOOK_SECRET }}" \
-H "Content-Type: application/json" \
-d '{"project":"ACME","sha":"${{ github.sha }}","status":"${{ job.status == 'success' && 'finished' || 'failed' }}"}'Migrations live in drizzle/ and run at server start. The history was squashed into a single 0000_baseline.sql generated from src/db/schema.ts. Its when in drizzle/meta/_journal.json is deliberately set to the timestamp of the last migration before the squash: the Drizzle migrator applies only migrations whose when is greater than the latest one recorded in the database, so existing databases (already at that migration) skip the baseline, and new databases create the schema from it.
Generate a new migration after editing the schema with npm run db:generate.
npm testAbout 280 Vitest tests run against a real Postgres (the one from docker compose); the service tests create their own tracker_test database from scratch. npm run typecheck runs tsc --noEmit.
- Single instance: timers, presence and realtime fan-out live in the process.
- Login: GitHub, Google or e-mail + password.
- The agent is Windows-first: the code is portable, but autostart exists only on Windows.
- The agent exe is unsigned, so Windows SmartScreen shows a warning.
- The UI is available in English and Polish.
- In the public demo, GitHub login, API tokens/MCP, the local agent, integrations and administration are disabled.
MIT, see LICENSE.