OpenKyrozen
A local-first terminal agent for coding, research, and project work, with durable memory and evidence-gated learning.
Install
The current published stable release is v2.0.8. It supports Python 3.12 and 3.13; Python 3.14 is outside the supported range.
macOS or Linux:
curl -fsSL https://raw.githubusercontent.com/EvanProgramming/OpenKyrozen/v2.0.8/install.sh | sh
kyrozen
Windows PowerShell:
irm https://raw.githubusercontent.com/EvanProgramming/OpenKyrozen/v2.0.8/install.ps1 | iex
kyrozen
On first launch, choose a provider and configure its credential. For example, set DEEPSEEK_API_KEY in the shell that starts OpenKyrozen. Read the installation guide before installing from a source checkout, updating, or recovering an interrupted update.
Try it
You: read this project and explain its architecture
You: investigate the failing tests and prepare a plan
You: accept the plan and fix the test
You: search for the latest Python release and cite the source
kyrozen opens the global workspace; kyrozen --project /path/to/repo binds a session to a project. The web interface is optional: run kyrozen-web --host 127.0.0.1 --port 8000 and visit http://127.0.0.1:8000. Configure KYROZEN_SERVER_TOKEN before binding beyond loopback.
The interaction modes make authority visible: Ask and Plan are read/network-only; Agent carries out explicitly requested or accepted work under capability and approval checks. These are application policy controls, not an operating-system sandbox.
What it does
- Operates on a selected workspace with file, shell, Git, web, browser, and GitHub tools.
- Supports hosted model providers, local Ollama, and named custom OpenAI-compatible endpoints, with simple/complex routing and configurable fallback.
- Persists sessions, events, claims, tasks, usage, and learning state in local SQLite. Optional vector and project graphs are derived indexes.
- Delegates suitable work to scoped sub-agents and collects reviewable results.
- Records outcome evidence and promotes bounded learning proposals only after validation. Learning does not fine-tune model weights or grant new capabilities.
- Provides terminal and optional web/API/MCP interfaces over the shared runtime.
Jev Decision is an optional judgment service for bounded routing and evidence checks. It may abstain or fall back; it never executes tools or approves work. Local Kev requires explicit consent. Compact prompts and tool discovery remain optional; the recorded pilot was inconclusive and the default profile is classic.
Documentation
The README is an overview. Detailed installation, command, provider, safety, storage, API, and operating instructions are maintained in the documentation index. Start with usage, configuration, or the security guide. The generated runtime inventory is the source of truth for tool names, HTTP routes, and MCP schemas.
Develop
Use Python 3.12 or 3.13. From a checkout, make install provisions the full development environment; make install-core selects the smaller non-browser path.
make check
make docs-check
make lint
make test
The development guide explains project structure, test profiles, packaging, acceptance checks, and release workflow. Contributions follow CONTRIBUTING.md; the project is licensed under MIT.