Kantaq
lionellau/kantaq
Issue tracker local para times pequenos, integrado com agentes IA
Ver no GitHubRastreador de tarefas local-first e agent-native para pequenos times que já executam agentes locais de IA (Claude Code, Cursor, Codex). Cada membro clona o repositório e executa em sua máquina; para times, aponta para um backend compartilhado (Supabase) que valida e armazena estado comprometido. Agentes conectam via gateway MCP loopback e propõem mudanças que humanos aprovam.
Instalação
npx convez-ai lionellau/kantaq1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Equipes de 2-10 pessoas gerenciando tarefas com suporte de agentes IA
- Desenvolvimento colaborativo com propostas de mudanças validadas por humanos
- Sincronização de estado de projeto entre múltiplos agentes locais
Veredito

Projeto ambicioso e bem-estruturado que resolve um gap real: sincronização local-first de tarefas entre agentes de IA e humanos, com segurança token-gated e aprovação explícita. Exige experiência técnica avançada, mas oferece modelo inovador para colaboração humano-agente em times pequenos.
README
A tradução para o português deste README ainda está na fila.
kantaq
Local-first, agent-native issue tracker for small teams (2–10) who already run a local AI agent (Claude Code, Cursor, Codex).
Each person clones the repo and runs the app on their own machine. A solo user runs it with zero backend. A team points every member's local copy at one shared sync backend (Supabase, later self-hosted Postgres) that holds committed team state. Agents connect only to each user's loopback MCP gateway and can propose changes; a human approves them.
There is no shared application instance and no server the project operates. The only shared thing is the sync backend — it just validates and stores committed events.
This repository holds the code. Product, architecture, and planning docs are maintained by the core team separately.
Status
v0.0.5 — the walking-skeleton core. On the Sprint 1 foundation (repo/env, data layer + migrations, local runtime + run modes, web shell, quality gates + shared test harness, identity + token-gated loopback auth), Sprint 2 adds the full hero loop: online sync through Supabase (E04/E24), the loopback MCP gateway with propose-first tools (E09/E10), and the screens — backlog, ticket page, Inbox, Members, My Agent (E19/E20/E21). End to end: create → sync → an agent proposes → a human approves → sync. See QUICKSTART.md.
Compatibility
kantaq supports three HTTP MCP clients — Claude Code, Cursor, and Codex — each with a generated connection snippet. The eight Tier-1 acceptance tests (first connection, role-aware read, propose + approve, permission denial, token rotation, untrusted-content tagging, session expiry, audit completeness) pass 8 / 8 against the official MCP SDK client in CI (make compat), and a real agent (Codex) connects, reads, and proposes end to end through the opt-in harness (make verify-agent). Per the badge rule (FR-E11-4), the Tier-1 (Reference) badge is advertised as certified only once a real client passes all eight at a pinned version — those runs are the release step, recorded with client version and date in the matrix:
→ docs/clients/compatibility.md — tiers, the 8 tests, client version, last-verified date, pass rate.
Connect your agent from Settings → My Agent: it generates the snippet for your own loopback gateway — .mcp.json (Claude Code), .cursor/mcp.json (Cursor), or ~/.codex/config.toml (Codex). See docs/mcp.md.
Quickstart
→ QUICKSTART.md — solo mode (zero backend) and team mode (shared Supabase, see docs/setup-supabase.md).
git clone https://github.com/lionellau/kantaq.git
cd kantaq
make setup # uv sync + pnpm install + build the web UI
make migrate # database migrations
make test # pytest + Vitest
make dev # FastAPI on http://127.0.0.1:3939 serving the built UI
A fresh clone reaches green (setup → migrate → test) in under 10 minutes (NFR-E01-1) — enforced by the fresh-clone CI gate.
Repository layout
kantaq/
pyproject.toml uv workspace + shared tool config + `kantaq` CLI
src/kantaq/ umbrella package: version + the `kantaq` CLI
apps/local-runtime/ FastAPI runtime: REST API + serves the UI (MOD-14)
packages/
protocol/ entities, canonical codec, Ed25519, grant verify (MOD-17)
sync_engine/ event log, snapshots, cursors, push/pull (MOD-04, MOD-26)
core/ tracker domain, resolver, recommendations, permissions
mcp/ MCP server, gateway checks, sessions, tools (MOD-08, MOD-09)
db/ SQLModel models + Alembic migrations (MOD-02)
web/ React + Vite SPA (MOD-10..13)
adapters/ sync backend adapters (Supabase MOD-05, self-hosted MOD-28)
evals/fixtures/ context-quality eval set (MOD-21)
docs/ stack ADR and code-side docs
.github/workflows/ CI gates
Dev commands
| Command | What it does |
|---|---|
make setup | install both toolchains and build the web UI |
make dev | run the FastAPI runtime on 127.0.0.1:3939 |
make migrate | run DB migrations |
make test | kantaq test → pytest + Vitest |
make lint | ruff + Biome |
make typecheck | mypy + tsc |
See docs/stack.md for the stack decision record (ADR-0001) and tool licenses.
Contributing
Read CONTRIBUTING.md. In short: conventional commits, the Golden rule (reuse before build), and every change ships with tests and an updated module spec.
License
Apache-2.0. See NOTICE.