Kryptic v1 is a local-first AI engineering agent CLI. It prioritizes a reliable workspace boundary, focused repository inspection, provider separation, direct command execution, bounded repair, verified skills, and evidence-based tests before any hosted interface is added. It runs on your own computer, uses your own provider keys, and does not require a hosted Kryptic account.
The fastest path for a fresh checkout is:
git clone https://github.com/adnqcr7-code/kryptic.git
cd kryptic
npm run v1:test
node src/cli.js startThe first interactive start asks for a provider, stores the key outside the workspace, checks local prerequisites, and opens chat. Run the complete local v1 release check with one command:
npm run v1:testFor everyday use, the concise commands are:
node src/cli.js chat
node src/cli.js setup
node src/cli.js "inspect the repository and improve the tests"
node src/cli.js fix "repair the failing test and verify it"
node src/cli.js test
node src/cli.js doctor
node src/cli.js history
npm run demoSee DEMO.md for the five-minute walkthrough and offline demo. See ROADMAP.md for measurable next milestones. See V1-AGENT-TEST.md for the agent-test contract and acceptance criteria. See CONTRIBUTING.md for safe contribution expectations and SECURITY.md for the prototype’s threat-model boundaries.
The project is an early local prototype. It currently provides:
| Capability | Status |
|---|---|
| Local CLI help | Available |
| Provider status for Google, OpenAI, and Claude | Available |
| Workspace inspection | Available |
| Focused repository search | Available |
| Recursive repository context inspection | Available |
| Path traversal protection | Available |
| Secret-looking file protection | Available |
| Guarded command policy | Available |
| Provider-backed planning | Available when a provider key is configured |
Bounded structured run command | Available when a provider key is configured |
| Exact-match patching and approval-gated edits | Available |
| Run-history audit records | Available |
| Provenance-aware local memory | Available |
| Offline demo without an API key | Available |
| Deterministic safety benchmark | Available |
| Interactive approvals | Available in TTY mode |
| Resumable run statuses and audit events | Available |
| Bounded plan–act–verify–repair loop | Available foundation |
| Transactional rollback and diff previews | Available |
| Interrupted-run resume | Available via resume |
| Strict response-size and criteria limits | Available |
Linked verify_workspace agent tool | Available |
| Structured verification evidence and failure classification | Available |
| Autonomous long-horizon coding loop | In progress |
Node.js 20 or newer is required. No npm dependencies are currently needed.
node src/cli.js --help
node src/cli.js providers
node src/cli.js inspect .
node src/cli.js search "TODO" --workspace .
node src/cli.js skills --workspace .
node src/cli.js skills "test repair" --workspace .
node src/cli.js ask "Explain this repository and plan a small test improvement" --provider openai --workspace .
node src/cli.js run "Inspect the repository and propose a safe test improvement" --provider openai --workspace .
node src/cli.js run "Fix the failing test and verify the repair" --provider openai --workspace . --approve --repair
node src/cli.js benchmark
node src/cli.js resume <run-id> plan.json --interactive --workspace .
npm testProvider credentials are read from environment variables and are never written by Kryptic into the repository. The supported variables are GEMINI_API_KEY, OPENAI_API_KEY, and ANTHROPIC_API_KEY. Model overrides are KRYPTIC_GOOGLE_MODEL, KRYPTIC_OPENAI_MODEL, and KRYPTIC_CLAUDE_MODEL.
Kryptic resolves paths against the selected workspace and rejects traversal outside it. It refuses secret-looking files such as .env, private-key files, and package-manager credential files. The command policy blocks destructive file operations, network-download commands, shell escapes, risky git operations, and unapproved commands. The policy is intentionally conservative and will expand only when tests justify the change.
The --repair run mode now connects provider-generated replacement plans to the bounded repair loop while preserving schema validation, approvals, rollback, and evidence-based success criteria. Verified procedural skills can be promoted only from status: verified evidence and discovered with skills. The model must never be allowed to invent a successful test result; verification comes from the local executor.
The most useful early contribution is a reproducible issue or a small fixture that reveals a safety, repair, provider, browser, or cross-platform failure. Please use the bug template or feature template. Testers can use the feedback form to report onboarding friction and structured evidence. Do not coordinate votes or post API keys.
The older Kairos repository is kept in a separate reference directory during development. Kryptic will reuse sound local-first and safety patterns without copying the entire Kairos control plane.
The Kryptic website is hosted separately from this repository: https://kryptic-site-meme-56ed.vercel.app