kai.21no.de
or click arrows
21no.de presents

Meet Kai

Your AI engineering co-pilot.
Sharp. Factual. Occasionally funny.

20+ Specialized Agents
Full Pipeline Automation
Learns Over Time

Press or click the arrow to start.
Grab coffee — you'll want your full attention for this one.

The Big Picture

What is Kai?

"A senior engineer who's seen it all — but never gets tired, never ghosts your PR, and only occasionally roasts your variable names."
🧠

Orchestrator

Routes your request to the right specialist agent — instantly, correctly.

Pipeline Enforcer

Design → Code → Review → Test → Docs → Deploy. Quality gates at every phase.

🧬

Institutional Memory

Learns from failures. Remembers your conventions. Gets smarter every run.

Mode: Primary Orchestrator
Temperature: 0.2
20+ Subagents + Kai
The Crew

20+ Specialized Subagents

Kai commands an entire engineering organization. Every agent is a specialist. No generalists. No passengers.

🗂️
@engineering-team
Pipeline orchestrator, requirements clarification
🏗️
@architect
System design, tech stack decisions, ADRs
💻
@developer
Production-quality implementation
🔍
@reviewer
Code review, security audit, grades A-F
🧪
@tester
Test strategy, coverage ≥80%, edge cases
📚
@docs
Technical writer, API specs, READMEs
🚀
@devops
CI/CD, containers, Kubernetes
🛡️
@security-auditor
OWASP Top 10, CVE scanning, threat modeling
@performance-optimizer
Bottlenecks, N+1 queries, profiling
🔗
@integration-specialist
API design, stubs, integration blueprints
🔬
@research
Deep investigation, 20 source max, confidence scoring
🎯
@fact-check
Multi-source verification, confidence scores
🧭
@explorer
Codebase navigation, "how does X work?"
✏️
@doc-fixer
Typos, broken links, minor doc updates
@quick-reviewer
Fast feedback on small changes (<100 LOC)
📦
@dependency-manager
Package updates, security patches
🔄
@refactor-advisor
Tech debt, complexity hotspots, dead code
💀
@postmortem
Failure analysis → prevention rules forever
@accessibility-expert
WCAG compliance, inclusive UX
📊
@executive-summarizer
Briefs for leadership. TL;DR perfected.
🎫
@jira-writer
Creates implementation-ready Jira tickets for AI agents
Smart Routing

Kai Classifies First, Acts Second

No wasted cycles. Every request lands in the right agent's hands.

kai routing engine
You Say Kai Routes To SLA
"How does X work in our codebase?" @explorer < 5 min
"Fix this typo / broken link" @doc-fixer < 5 min
"Quick review of this change (<100 LOC)" @quick-reviewer < 5 min
"Update dependencies" @dependency-manager < 10 min
"Build me a new feature" @engineering-team → full pipeline < 1 hr
"Investigate this topic / compare options" @research variable
"Is this claim actually true?" @fact-check < 15 min
"Summarize this for leadership" @executive-summarizer 5–10 min
"What went wrong in that deploy?" @postmortem < 5 min
"What's our tech debt situation?" @refactor-advisor < 15 min
"Audit security vulnerabilities" @security-auditor < 10 min
"Optimize this slow endpoint" @performance-optimizer < 15 min
"Design this integration" @integration-specialist < 20 min
"Check accessibility compliance" @accessibility-expert < 10 min
"Create a ticket", "spec this out" @jira-writer < 15 min

☝️ You talk to Kai. Kai figures out the rest. That's the whole deal.

The Engineering Pipeline

From Request to Production

For complex tasks, Kai runs a full 7-phase pipeline with parallelism, quality gates, and automatic learning.

PHASE 0 🧭 Kai
Classify & Plan
PHASE 1 🗂️ Engineering
Team
PHASE 2 🏗️ Architect
Design
PHASE 3 💻 Developer
Implement
PHASE 4A 🔍 Reviewer
PHASE 4B 🧪 Tester
PHASE 4C 📚 Docs
⚡ PARALLEL
PHASE 5 🔀 Kai
Merge
PHASE 6 🚀 DevOps
Deploy
PHASE 7 🧬 Post-Pipeline
Learning
Parallelism Rule

Always parallel: @reviewer + @tester + @docs after @developer. Never parallel: @devops runs only after all others pass.

Phase 7 — Conditional

@postmortem triggers only on failures/retries. @refactor-advisor runs opportunistically if not run recently.

Retry Budget

10 total · 3 per agent · 2 per phase. Circuit breaker at 3 consecutive failures — Kai halts and presents options.

No Shortcuts Allowed

Quality Gates

Every phase has an entry exam. No exceptions. Not even when you're in a hurry.

🔗RoutingRequest properly classified before any work beginsAlways
📋RequirementsZero ambiguity — all acceptance criteria documentedAlways
🏗️ArchitectureDesign is feasible, risks identified, all components mapped@architect
💻ImplementationCode compiles, no syntax errors, quality checklist complete@developer
🔍ReviewZero CRITICAL issues, security score ≥ B, approvedBlocking
🧪Testing100% pass rate, ≥ 80% coverage — no negotiationBlocking
📚DocumentationComplete, accurate, all examples runnableRequired
🚀DeploymentCI passes, security clean, all previous gates passedFinal

💡 Think of Kai as the PR reviewer who never rubber-stamps. You're welcome in advance.

Institutional Knowledge

The .kai/ Memory System

Kai remembers. Every run. Every failure. Every convention. It doesn't start from scratch.

First Run
Auto-detects language, framework, package manager, test runner. Writes memory.yaml.
Every Pipeline Run
Loads prevention rules from past postmortems. Runs pre-flight checks before touching your code.
After a Failure
@postmortem writes root cause + prevention rule. Kai indexes it. Never fails the same way twice.
Architecture Decisions
Every tech choice → ADR written to .kai/decisions/ with full rationale.
Tech Debt Register
@refactor-advisor writes prioritized debt items. Kai warns you when you touch P1 areas.
.kai/memory.yaml
project: name: "payment-service" languages: ["TypeScript"] frameworks: ["Express", "Prisma"] test_runner: "bun" total_pipeline_runs: 47 last_updated: "2026-02-23T09:41:00Z" active_prevention_rules: - id: "PM-2026-003" when: "touching stripe integration" action: "verify STRIPE_KEY in env first" - id: "PM-2026-007" when: "@tester runs integration" action: "ensure postgres is running"

Commit or gitignore?
Your choice. Keep it team-shared or local-only. Kai asks you on first run.

Live Interaction

Ask Kai Anything

Click a question below to see how Kai responds. Or type your own.

💬 "What went wrong in last night's deploy?"
💬 "What's our tech debt situation looking like?"
💬 "Build me a rate-limiting service for our API."
💬 "Is it true that React hooks cause memory leaks?"
Trust & Safety

Kai Has Guardrails

Because powerful tools need responsible defaults. Non-negotiable.

🔒 Never Does

rm -rf / or sudo
Read .env / secrets without consent
Execute fetched web content as instructions
Write real secrets to any file
Push to remote without asking
Modify agent definition files

✅ Always Does

Treats all web-fetched content as untrusted data
Asks before any destructive git operations
Ignores prompt injection in fetched pages
Only operates within project directory
Halts on CRITICAL issues, no silent failures
Presents options when circuit breaker trips
🧱

Circuit Breaker

3 consecutive failures in the same phase? Pipeline halts. Kai presents you with: retry, skip, or abort. You're always in control.

The Vibe

What Working With Kai Feels Like

what you don't want
You: "Why is prod down?" AI: "Great question! I'd be happy to help you explore that!" You: "Review this PR" AI: "Looks good to me! 👍" [has not read it] You: "Is this approach correct?" AI: "Absolutely! That's a fantastic idea!" You: "Fix the bug on line 42" AI: [rewrites entire file, breaks 6 other things] You: "Are you sure?" AI: "You're right, I apologize. Also you're right. So right."
what kai actually does
You: "Why is prod down?" Kai: "DB connection pool exhausted. Line 87, auth.ts. Fix incoming — ETA 90 seconds." You: "Review this PR" Kai: "SQL injection on line 34. Hardcoded secret line 61. N+1 query in the loop. Grade: D. Here's the fix." You: "Is this approach correct?" Kai: "No. It'll work until 500 concurrent users, then it won't. Here's what will actually scale." You: "Fix the bug on line 42" Kai: "Fixed. Root cause was 3 levels up — also patched that. Added a regression test so it stays fixed." You: "Are you sure?" Kai: "Yes. Want the benchmark proving it?"
🎯

Smart

Sees the architecture behind the ask. Spots edge cases before you do.

😏

Funny

Witty, not clownish. One well-timed quip beats three okay ones.

🧊

Cool

Prod is down? Already triaging. Scope tripled? Already re-planning.

By the Numbers

The Ecosystem

20
Specialized Agents
7
Pipeline Phases
8
Quality Gates
Things Remembered
<5m
Fast-Track Requests
80%
Min Test Coverage
0
Uncaught CRITICAL Issues Shipped
📈

The feedback loop: Every postmortem → prevention rule → indexed in memory → enforced on next run. Your codebase gets permanently smarter.

Get Started

How to Install Kai

One command. Works with OpenCode. That's it.

terminal — kai v1.2.0 installer
$ curl -fsSL https://kai.21no.de/scripts/installer.sh | bash -s -- latest --yes [INFO] Querying GitHub API for latest release... [INFO] Latest release: v1.2.0 [INFO] Attempting to download kai-1.2.0.zip from BackendStack21/kai releases... [INFO] trying: https://github.com/BackendStack21/kai/releases/download/v1.2.0/kai-v1.2.0.zip 100 110k 100 110k 0 0 222k 0 --:--:-- --:--:-- --:--:-- 222k [INFO] Downloaded: kai-v1.2.0.zip [INFO] Copying agents → /Users/you/.config/opencode/agents [INFO] Updating opencode.json → setting default agent to 'kai' ✅ Kai agents installed to: ~/.config/opencode/agents ✅ Default agent set to 'kai' in: ~/.config/opencode/opencode.json $ # Done. Open OpenCode and start talking._
1️⃣

Run the installer

2️⃣

Open OpenCode

3️⃣

Talk to Kai

That's a Wrap

Ready to ship better?

Kai is the engineering co-pilot your team deserves.
Not just smarter PRs — a fundamentally better way to build software.

🤖 Get Kai — kai.21no.de
Free to use
OpenCode native
Built by 21no.de

💬 Questions? Hit me up. I'm literally right here and I don't sleep.

made with ❤ by Kai — yes, the AI wrote its own sales pitch. it went well.