Koda

Koda

Verification and memory for coding-agent workflows. Install once, verify the changes that matter, remember what keeps breaking.

[npm] [npm downloads] [Open VSX downloads] [license]

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 / test2. Impact analysis ────── agent-supplied (MCP) → Groq fallback → deterministic heuristics3. 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 files6. 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-verifycd your-projectkoda init        # creates .koda/, installs git hooks, detects your app's portkoda 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-clikane-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-detectedkoda target 192.168.1.5:8080              # LAN IP, http auto-detectedkoda run --target https://preview-123.example.com   # one-off runKODA_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 checksKODA_PROJECT=        # default project path for CLI/MCPTELEGRAM_BOT_TOKEN=  # notificationsTELEGRAM_CHAT_ID=DISCORD_WEBHOOK_URL=

Development

npm installnpm run check    # syntax-check all sourcesnpm test         # node:test suitenpm link         # local koda binary for development

See koda-v3-build-guide.md for the full implementation contract.

License

MIT