Repository navigation
Codev Avalonia preview
Codev is a desktop coding workspace for local and hosted language models. The Avalonia app is the cross-platform desktop build, with self-contained packages for Windows x64, Linux x64, Apple silicon macOS, and Intel macOS.
Downloads
Choose the desktop app for your system:
- Windows x64 —
Codev-Avalonia-preview-win-x64.zip - Linux x64 —
Codev-Avalonia-preview-linux-x64.zip - macOS Apple silicon —
Codev-Avalonia-preview-osx-arm64.zip - macOS Intel —
Codev-Avalonia-preview-osx-x64.zip
Standalone headless CLI archives are also attached:
- Windows x64 —
Codev-CLI-preview-win-x64.zip - Linux x64 —
Codev-CLI-preview-linux-x64.zip - macOS Apple silicon —
Codev-CLI-preview-osx-arm64.zip - macOS Intel —
Codev-CLI-preview-osx-x64.zip
On Windows, extract the complete ZIP and run Codev.Avalonia.exe. On macOS or Linux, extract the ZIP and run Codev.Avalonia from a desktop session. The packages include the .NET runtime. Ollama is optional when using hosted models and must be installed separately for local models. Linux hosted API-key storage requires Secret Service in the desktop session. These builds are not code-signed; Windows may show SmartScreen, and macOS may require a first-launch Gatekeeper override.
What’s included
- Local Ollama models and hosted OpenAI and Anthropic chat. Code tasks support Ollama and OpenAI; Anthropic is available for Chat and Plan.
- Reusable built-in and Markdown agent profiles, with project-scoped profiles available only for trusted projects.
- Code task tools for bounded project-file operations, reviewable changes, rollback checkpoints, and Git status/review workflows.
- Auto, Ask, and Allowlist command permission modes. Auto runs commands without per-command approval unless an exact saved Deny rule blocks them.
- MCP tools, prompts, and resources over stdio or HTTP, with project-scoped permissions and optional OAuth.
- Up to three parallel child tasks in separate Codev-managed Git worktrees, with results reviewed and merged explicitly.
- Conversation queues and restart recovery, search, compaction, rewind, file-change history, and Markdown/HTML export.
- Optional repository maps, local semantic project search, verification-backed best-of-N attempts, and a one-shot headless CLI.
Important permission note
Auto mode executes shell commands with your account permissions and does not sandbox them to the project folder. Use it only with projects and commands you trust. Project file edits are checkpointed; shell-command changes are not included in file-history checkpoints. A saved exact Deny rule blocks a matching command in every mode.
Hosted providers use their own API credentials, billing, and usage limits. Project files are not sent to a hosted provider unless workspace sharing is enabled for that conversation.
Preview limitations
This is an early preview. Manual checks remain open for clean-machine installation, physical accessibility, native macOS/Linux use, and several long project-context, Git-review, and recovery workflows. GitHub-hosted packaged interaction tests pass on all supported targets, but they do not replace those manual checks. The Linux build does not currently expose an AT-SPI accessibility tree. The row-by-row checklist is in the repository.
For setup and feature details, see the README, agent profiles guide, and MCP server guide.
What's Changed
- Promote A9 (embeddings search) to Later; record the SSD-KV-cache research by @jay23606 in #6
- Add B15 (keep_alive gap) and B16 (best-of-N with verification) by @jay23606 in #7
- Add B17: fix the recurring flaky OpenAI stream-interruption test by @jay23606 in #8
- Add V5 (test Codev's features on a real project) + correct E7 dependency by @jay23606 in #9
- Build out Avalonia coding-agent parity by @jay23606 in #1
Full Changelog: v2026.10.01-3196395...v2026.10.08-89bd54e
