Koda
Verification and memory for coding-agent workflows. Install once, verify the changes that matter, remember what keeps breaking.
1,000+ downloads on Open VSX — Koda is already watching commits in editors around the world.
Why Koda exists
Coding agents write features in minutes — but who verifies every commit? Full test suites are too slow to run per commit, so most verification happens late (or never). Koda closes that gap: on every commit it verifies only what changed, in seconds-to-minutes, and remembers what keeps breaking so your agent stops reintroducing old bugs.
Koda never writes code. It orchestrates verification, remembers patterns, and reports — to both the human and the agent.
The three surfaces
How it works — the pipeline
git commit
│
▼
pre-commit ── secret guard (blocks committed API keys / private keys)
│
▼
post-commit ── koda run (background, never blocks you)
│
▼
1. Read diff ──────────── classify files: api / page / component / lib / config / test
2. Impact analysis ────── agent-supplied (MCP) → Groq fallback → deterministic heuristics
3. Kane browser flows ─── headless, scoped to changed UI only (page/component changes)
4. Endpoint checks ────── GET / HEAD / OPTIONS only; regressions flagged (200 → 500)
5. Integration tests ──── Jest / Vitest / Mocha / node:test, matched to changed files
6. Report ─────────────── JSON (agent-readable) + Markdown (human-readable)
7. Memory ─────────────── recurring patterns, resolution history, time-to-fix
Most commits never launch a browser: API/config/test-only changes verify with endpoints + tests in seconds. Browser flows fire only when UI files change.
Quick start
npm install -g koda-verify
cd your-project
koda init # creates .koda/, installs git hooks, detects your app's port
koda doctor # verifies: git, config, Kane CLI, target URL, directories
Then just keep committing — verification runs in the background on every commit.
One-time browser setup (optional, per user)
Browser flows are powered by the Kane CLI. The IDE extension onboards it for you: click Enable and Koda offers to install Kane and connect your account. Or do it manually:
npm install -g @testmuai/kane-cli
kane-cli login --oauth # opens your browser — use (or create) YOUR OWN Kane account
About Kane accounts — yours, not ours
Koda ships zero credentials. Every Koda user authenticates their own Kane account:
- The extension's OAuth flow opens the Kane login in your browser; new users register there
- Your session lives locally in
~/.testmuai/on your machine — never synced, never proxied - Browser-flow credits come from your Kane account, not the developer's
- In CI, each repo configures its own
KANE_USERNAME/KODA_ACCESS_KEY-style secrets
No Kane? No problem. Koda runs in degraded mode: endpoint checks, integration tests, memory, and reports all work; browser flows are honestly marked SKIPPED — never faked.
Target URLs — localhost, LAN, or hosted
Koda verifies whatever your app points at. The default is auto-detected at koda init (probes common dev ports: 3000, 5173, 8080, 8000, 4200, 5000…). Override it three ways:
koda target https://myapp.vercel.app # permanent — hosted URL, https auto-detected
koda target 192.168.1.5:8080 # LAN IP, http auto-detected
koda run --target https://preview-123.example.com # one-off run
KODA_TARGET=https://staging.example.com # per-session / CI via env
Precedence: --target flag → KODA_TARGET env → .koda/config.json → auto-detected default.
CLI reference
Agent integration (MCP)
claude mcp add koda -- npx koda-verify mcp
Works with any MCP client: Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Zed. In a development checkout: claude mcp add koda node /absolute/path/to/koda/mcp/server.js.
MCP tools
The agent loop
agent edits code → commits → hook fires koda run → report written
→ agent calls koda_verify (its own flows) / koda_report via MCP
→ reads the verdict → fixes what failed → commits again
→ memory marks the failure RESOLVED (with time-to-fix recorded)
Reports record analysis_source: 'agent' when the agent supplied flows — verifiable proof the loop is agent-driven.
Reports
Every run writes a timestamped pair to .koda/reports/:
.koda/reports/2026-08-30T14-35-28-909Z-9416b29.json ← agent-readable
.koda/reports/2026-08-30T14-35-28-909Z-9416b29.md ← human-readable
The JSON includes: commit metadata, analysis_source (agent / llm-fallback / fallback), risk level, per-check results with evidence file paths, memory context (similar failures, patterns, fix confidence), and action_required with priority and suggested focus. Markdown renders PASS / FAIL / SKIPPED distinctly with coverage status.
Raw Kane session events are preserved per-flow under .koda/evidence/.
Memory — remembers what keeps breaking
- Every failed check gets a stable key; failures that recur 2+ times become patterns
- When a fix lands and the check passes, the failure is marked resolved with
time_to_fix_ms - Coverage-gap findings ("no tests cover
app/page.js") resolve automatically once a passing test covers those files - Skipped checks (app offline) never pollute memory
- Writes are atomic; corrupt memory files are quarantined, not deleted
Safety guarantees
- Non-destructive hooks — existing pre-commit/post-commit/pre-push hooks are backed up before appending; Husky respected; idempotent via marker
- Read-only HTTP by default — only GET / HEAD / OPTIONS; mutating methods require explicit opt-in (
allow_mutating_methods: true) plus agent-declared endpoints - Honest results — if the app is down, checks are
SKIPPED, never reported as failures; skipped checks never enter memory - No secrets shipped — the package contains zero credentials; the pre-commit guard blocks accidentally committed keys
CI/CD
koda cicd
Generates .github/workflows/koda.yml from a validated template: Node setup, npm ci, build-if-present, Kane CLI install + auth via GitHub secrets (KANE_USERNAME, KANE_ACCESS_KEY), app start, wait-on target, koda run --commit ${{ github.sha }}. Template-first — no LLM-generated YAML. Refuses to overwrite without --force.
Configuration reference (.koda/config.json)
{
"kane": {
"target": "http://localhost:3000",
"headless": true,
"max_steps": 30,
"timeout_ms": 300000
},
"tests": {
"browser_flows": true,
"endpoint_tests": true,
"integration_tests": true,
"allow_mutating_methods": false,
"expected_statuses": { "GET /api/health": [200] }
},
"notify": { "telegram": false, "discord": false }
}
Environment variables
Optional — Koda works without any of these. Loaded from the project .env, falling back to ~/.koda/.env (configure once, use everywhere):
GROQ_API_KEY= # Groq fallback impact analysis (agent runs skip this entirely)
KODA_TARGET= # override target URL (flag > env > config)
KODA_TEST_TOKEN= # bearer token for authenticated endpoint checks
KODA_PROJECT= # default project path for CLI/MCP
TELEGRAM_BOT_TOKEN= # notifications
TELEGRAM_CHAT_ID=
DISCORD_WEBHOOK_URL=
Development
npm install
npm run check # syntax-check all sources
npm test # node:test suite
npm link # local koda binary for development
See koda-v3-build-guide.md for the full implementation contract.
License
MIT