stk/

The drift‑proof framework for AI‑built apps.

stk is a strict Python framework for AI coding agents. Every session lands in the same architecture, passes the same gate, and proves the app still works in a real browser, so what agents build stays reliable, secure, and easy to change.

paste into your coding agent

Build me an app on stk. Read stkit.dev/start.md and follow it step by step. The app: [describe your app]

your agent, in ~/acme
$ uvx --from stkit stk init acme
+ created acme/acme/app.py
+ created acme/CLAUDE.md and five agent skills
+ created acme/.env (fresh secrets)
$ uv run stk new invoice
+ added invoice to the extensions in acme/app.py
+ migration generated and applied for invoice
$ uv run stk verify
✓ ruff
✓ checks
✓ migration-drift
$ uv run stk smoke
PASSED dashboard /dashboard/
PASSED /invoices/ /invoices/ .stk/smoke/invoices.png
PASSED /invoices/ dark /invoices/
Smoke status: passed

Typing was never the hard part.

Any framework will let an agent write code. The trouble is everything after: knowing what the app actually is, where a change belongs, what just broke, and whether the page renders at all. That part is stk.

the whole loop, with real output →

Less guessing, more asking

Instead of guessing from files, the agent asks. stk doctor answers whether the app can start, and every line that isn't a tick carries the command that fixes it. stk inspect reports routes, models and config as JSON, with auth read from the real guard chain.

$ stk doctor
✓ migrations     at heads (d7e01db94a41, stk_0008)
✓ admin user     1 admin(s)
! browser        chromium not installed
    → uv run playwright install chromium
$ stk inspect routes --json   # condensed
GET,POST  /tf-setup   auth  security.two_factor_setup
GET       /login      open  security.login

A place for everything. Agents included.

stk new writes a feature as an extension inside your app, with its model, views, template, seed rows and an applied migration, and hands back the URL. The agent never edits a framework file, so upgrading stk is a version bump.

$ stk new invoice
  + created acme/invoice/models.py
  + created acme/seed.py with sample invoice rows
  + added invoice to the extensions in acme/app.py
  + migration generated and applied for invoice
Open: http://localhost:5000/invoices/

Every error ships with its own fix

stk verify runs lint, the framework's own checks, and a migration drift test against a throwaway database. Each failure prints the command that clears it, in --json too, so the agent applies the fix instead of inventing one. A check that could not run fails.

$ stk verify
✗ migration-drift
    add_column: invoice due_date
    → uv run stk db revision -m "describe change" && uv run stk db upgrade
✓ checks

A 200 is not proof

stk smoke opens every page in Chromium, light and dark, on a throwaway database filled by your seed.py, and saves a screenshot of each. A page that answered 200 but never rendered fails. So does invisible text, a raw id or undefined on screen, and a check that could not run.

PASSED /invoices/       HTTP: 200  screenshot: .stk/smoke/invoices.png
PASSED /invoices/ dark  HTTP: 200  screenshot: .stk/smoke/invoices-dark.png

Briefed before the first prompt

Every new app ships with the rules an agent would otherwise learn by breaking things: a CLAUDE.md (also read as AGENTS.md), five skills for blueprints, APIs, migrations, deploys, and lessons, and a CI workflow that runs the tests, the verify gate, and a real browser on every push.

CLAUDE.md  → AGENTS.md
.claude/skills/  stk-api stk-blueprint stk-deploy stk-lessons stk-migrate
.github/workflows/checks.yml  tests · stk verify · stk smoke

Four bugs the tests waved through.

All in stk itself, all invisible to a green test suite and a clean console. The hound found them anyway. Each one is now a check that can’t pass quietly again.

route auth
Seven security routes, including 2FA setup and passkey registration, were reported as public while actually requiring auth. An agent asking what is exposed was pointed at the wrong ones.
contrast
A primary button rendered invisible at 1.0:1 after a cascade-layer change in Vuetify 4. No console error and no failed request.
real browser
The dashboard hung forever on a production Vue build. The development build only warned, and the app had been shipping the development build.
page sweep
GET /mf-recovery returned 500 to anonymous users. A security route, unauthenticated, broken in production.

The parts you don’t want an agent improvising.

Auth, jobs, migrations: the expensive-to-get-wrong parts are already built, tested, and boring on purpose. They live in the framework, not in generated code. Dependencies are Python only, and your users stay in your own database.

auth
Login, server-side sessions, and password reset by email, plus TOTP 2FA, WebAuthn passkeys (Touch ID, YubiKey), OAuth (Google, GitHub), recovery codes, account lockout, and rate limiting. Public sign-up is off by default, and confirms the email when you turn it on.
2FA · TOTPWebAuthnOAuthArgon2id
admin
User CRUD, role-based access, an activity audit log, and a jobs page, on paginated tables with a JSON API.
database
Async SQLAlchemy on SQLite by default, PostgreSQL when you need it, with Alembic migrations.
jobs
Durable jobs that survive a restart: register a function, queue it, run stk worker. A failure rolls back, keeps its traceback, and retries. No Celery, no Redis.
realtime
Authenticated, origin-checked WebSockets and in-app notifications that reach a user's open pages from any process: a web worker, a job, or the shell.
files
Opt-in uploads stored once by their sha256, typed from the bytes, size and type limited, with WebP thumbnails for images.
agent tools
Expose your own functions to agents over MCP with one decorator. Personal tokens are stored hashed, expire, and carry only the user's current roles.
ui
Vue 3 + Vuetify 4 loaded from pinned static files, with 5,000+ Tabler icons. No JavaScript build step, so the agent edits plain HTML templates.

One prompt.
A working app.
Receipts attached.

Your agent installs stk, writes the app with fresh secrets, creates the admin user, and proves it green with doctor, verify, and a real-browser smoke test before it builds anything. The steps it follows are public, so you can read them first.

paste into your coding agent

Build me an app on stk. Read stkit.dev/start.md and follow it step by step. The app: [describe your app]

Ship it with the built-in deploy script (Caddy with automatic TLS, a firewall, fail2ban, key-only SSH), with Docker Compose, or as plain ASGI behind your own proxy. deploy docs

Give your agent the docs.

A read-only MCP server answers from the same pages as docs.stkit.dev, with three tools: search_docs, get_page, and get_command. No sign-in.

mcp.stkit.dev
# Claude Code
$ claude mcp add --transport http stk-docs https://mcp.stkit.dev/mcp
# Codex
$ codex mcp add stk-docs --url https://mcp.stkit.dev/mcp
# Cursor, ChatGPT, Claude, anything else: a custom MCP server, no auth
https://mcp.stkit.dev/mcp

Built with stk

Sealbook

Evidence intelligence for defensible investigations. AI proposes claims pinned to the exact line, frame or region they came from, people decide with a passkey, and findings export as a signed package anyone can verify offline.

sealbook.org →
Ziglag

Open-source, self-hosted invoicing for freelancers and small businesses. No document limits.

view Ziglag →
Qarina

An AI research assistant for human rights investigators. Deterministic collection lanes, content credentials, an archived citation trail and a hashed report. AI finds leads. Humans establish facts.

qarina.org →