Reflexor
Local-first AI coding assistant for VS Code and Cursor. Chat your codebase, complete inline, retrieve with local RAG, and keep Agent writes behind your approval.
Install for VS Code | Install for Cursor
Why developers use Reflexor
- Codebase chat: Ask about the current file, a selection, files you attach, and indexed workspace context.
- Local RAG: Search the project with embeddings that stay on your machine.
- Inline completion: Ghost suggestions while you type, without sending the buffer to a Reflexor cloud.
- Approval-first Agent: Propose edits and commands, then wait for your review. Agent stays experimental.
- Hardware-aware setup: Reflexor recommends a chat and embedding pack for your machine. You confirm before any large download.
- Optional web research: Keep the default local workflow, or allow Web Search when a task needs current public information.
Install and first run
- Open Reflexor Chat.
- If the local engine is missing, choose Set up local engine and confirm the one-time download.
- When Reflexor recommends models, choose Continue and download. Model weights are never downloaded silently.
- Use Ask for a question, Plan for a written approach, or Agent to propose an edit. Review any change before it lands.
You do not need to learn model names or configure a terminal. Chat works after the engine and a model are in place. Workspace indexing is optional.
Before you install
- Open only trusted local folders. Untrusted and virtual workspaces are not supported.
- The VSIX does not ship the engine or model weights. Those downloads happen only after you confirm.
- CPU-only machines are supported. Speed depends on the model and your hardware.
- Local chat, completion, and indexing do not need a cloud account.
- Agent can propose file writes and shell commands. Those stay experimental and need your approval.
Reflexor is free to use and proprietary. Source code is maintained privately and is not distributed under an open-source license.
- Visual Studio Code
- Cursor through this Open VSX listing
- Windows x64
- Linux x64
- macOS Apple Silicon (
darwin-arm64)
The listing shows the editor version it needs. The current package requires VS Code ^1.106.0 or a compatible Cursor build.
linux-arm64 and macOS Intel are not in this package. Install the VSIX that matches your operating system and architecture.
Models and local engine
Reflexor can:
- reuse a system Ollama installation when one is already available
- download and manage its embedded local engine after you confirm
- recommend chat and embedding models based on available hardware
- use local embeddings for richer codebase retrieval
The VSIX does not ship model weights or an Ollama binary. Those downloads happen only after confirmation.
Model licenses belong to their providers, not to Reflexor. Speed and quality depend on the selected model, CPU, memory, optional GPU, and project size. CPU-only machines are a supported mode.
Chat, context, and Agent
Use Ask for explanations, Plan for a written approach, and Agent for experimental workflows that propose edits or commands.
Reflexor can use:
- the selected code
- the active file
- files you explicitly add to chat
- local workspace indexing and retrieval
- diagnostics and project structure when the workflow needs them
Review the proposed diff before accepting it. Agent workflows are experimental and can still produce incorrect plans or code. Session auto-approve is off unless you turn it on. Policy, path, and revalidation checks still apply.
Web Search, Web Research, and remote inference
The core product is local-first. Optional network features stay off until you allow them.
- The first explicit web request shows the sanitized query and asks for permission (once, session, or cancel).
- Persistent enablement lives in Reflexor Settings.
- Web Search sends search query text. It does not send code, file contents, absolute paths, or RAG context.
- Private or localhost URLs require a separate approval.
- Pointing inference at a non-local Ollama endpoint requires a separate consent. Prompts and embeddings then go to that endpoint, not to Reflexor.
Privacy
See the full privacy policy.
- Core chat, inline completion, project indexing, and local embeddings run on your computer.
- Conversation history and prompts stay local for local-engine turns.
- Product telemetry is off by default and, when enabled, remains on your machine.
- You can export or delete local conversations from the Command Palette.
Reflexor does not claim to be fully offline or private under every configuration. Read the privacy policy before enabling web features, remote inference, or integrations.
Requirements
- Visual Studio Code or Cursor that meets the version shown on the listing
- Windows x64, Linux x64, or macOS Apple Silicon
- A chat or coder model, installed only after confirmation
- Optional embedding model such as
nomic-embed-text for richer workspace retrieval
- Linux:
tar with zstd support, or a zstd binary, for the embedded engine archive
- macOS: Apple Silicon (
darwin-arm64). Metal-capable hardware is recommended when available
Troubleshooting
- Backend offline: Reload the window and check Output: Reflexor.
- Local engine offline: Wait a few seconds, then reopen Reflexor Chat and use Set up local engine.
- No models: Use Continue and download in the Reflexor setup checklist.
- Slow completion: Choose a smaller model and avoid oversized context overrides.
- Stale project context: Rebuild the local project index after large refactors.
- Need a bug report: Run Reflexor: System Diagnostics, redact secrets, and email the excerpt.
Agent workflows are experimental. Performance varies with model, hardware, context size, and project size.
Support and legal
Feedback after a completed workflow is welcome, whether the result was good or not. Reflexor never requires a rating or a positive review.
License
Reflexor is proprietary software. The package license and legal terms are described on the Reflexor terms page. The complete license text is included in the installed package.
Third-party components remain under their own licenses and notices included in the installed package.
Source code is maintained privately and is not open source.
