@kognai/build

The Kognai sovereign orchestrator as a CLI. Submit a task or a goal; get reviewed, compliance-gated deliverables — or a rejection with reasons. Runs on the cloud or fully local ($0).

npx @kognai/build "TypeScript email validator function"

What it does

Every task runs a real pipeline, not a single LLM call:

generate → code review (primary) → code review (second pass) → compliance review → ship | reject

The third review (compliance) checks constitutional, legal, and security concerns and is non-overridable — a code-grade-A deliverable with a compliance FAIL is rejected. That's what distinguishes the Kognai runtime from a thin LLM wrapper.

Under the hood it routes every model call through @kognai/orchestrator-core, so you get tiered routing, cost accounting, and an optional local ($0) backend for free.

Install / Run

# zero-install
export ANTHROPIC_API_KEY=sk-ant-...
npx @kognai/build "<task>"

# or globally — gives you `kognai` (REPL) and `kognai-build` (one-shot)
npm install -g @kognai/build
kognai            # interactive REPL
kognai-build "<task>"   # one-shot

Modes

# Task — one deliverable through the triple-supervisor
kognai-build "Stripe webhook handler" --out src/webhook.ts

# Goal — decompose into tasks, run each (TICKET-135 hierarchy)
kognai-build --goal "Express CRUD service with Postgres" --out-dir ./svc

# Sovereign — all inference local via Ollama ($0, no API key)
kognai-build "log shipper" --sovereign

# Swarm — delegate to the FULL orchestrator-core swarm
#   (CEO planning · CTO governance gate · dual-supervisor · reconciliation)
kognai-build --swarm "auth service with JWT" --out-dir ./auth

# Mandate — sign a PACT mandate (cost envelope) before executing
kognai-build --mandate "research synthesis tool" --out-dir ./tool

# Optimization profile — coder + reviewers optimize for one axis
kognai-build --goal "payments API" --profile security

Flags: --out <path> · --out-dir <dir> · --sovereign · --swarm · --mandate · --profile <axis> · --no-compliance

Sovereign mode

--sovereign routes all inference to a local Ollama instance — coder, supervisors, and (in --swarm) the CEO/CTO too. No ANTHROPIC_API_KEY required, $0 cost floor. Requires Ollama running locally with a capable model (e.g. qwen3:14b).

Workspace (.kognai/)

Drop a .kognai/config.yaml in your project and the CLI binds to it: deliverables go to your configured root, runs are measured (KSL), and you can extend the registries via .kognai/tools.json / .kognai/skills.json.

Configuration

| Env var | Default | Purpose | |----------------------|----------------------|-------------------------------------------| | ANTHROPIC_API_KEY | (required for cloud) | Claude API auth — not needed with --sovereign | | CLAUDE_MODEL | claude-sonnet-4-6 | Override the cloud coder + reviewer model |

Compliance gate

The non-overridable third reviewer checks:

Constitutional (the 3 Laws): 1) Never Harm · 2) Earn Existence (no fake/deceptive functionality, fabricated metrics, hidden surveillance) · 3) Transparency to Creator (no obfuscation, embedded credentials, undisclosed telemetry).

Legal / Regulatory: GDPR consent gates · tax-evasion shortcuts · copyleft lift.

Security (auto-fail): hard-coded secrets · injection vectors · disabled TLS · eval on untrusted input.

A FAIL on any category rejects the deliverable. Override with --no-compliance only for trusted local prototypes — never production.

Cost

Cloud: ~$0.01–0.05 per task (the footer prints actual spend). --sovereign: $0. --swarm runs the full leadership ceremony and costs more per goal.

License

MIT