Reflexor
Local AI coding assistance
Reflexor is a local-first AI coding assistant for Visual Studio Code and Cursor on Windows x64. It helps you complete code, chat about your project, retrieve workspace context, and run experimental Agent workflows — with inference through a local Ollama installation.
Reflexor is free to use, but it is proprietary software. Its source code is maintained privately and is not distributed under an open-source license.
Version shown: Reflexor 0.1.0-alpha.1
Package version: 0.1.0
Channel: Visual Studio Marketplace Pre-Release (when published)
Platform: Windows x64
Features
- Inline completion (Ghost) — local fill-in-the-middle style suggestions while you type
- Chat — Ask / Plan / Agent modes with a code-aware assistant
- Workspace context — indexing and retrieval so answers can use your project files
- Request Understanding — routes intent toward the right workflow
- Experimental Agent — propose edits and commands with approval gates
- Hardware-aware model selection — including a qualified 4 GB operating policy when enabled
- Web Search / Web Research — optional web-assisted answers (on by default in this alpha)
How it works
- The extension starts a bundled local backend on your machine.
- Models run through Ollama on localhost.
- Chat, completion, and indexing stay on your computer unless you use web features.
- Agent file writes and shell runs require approval unless you explicitly change that behavior.
Requirements
- Windows x64
- Visual Studio Code ≥ 1.80 or Cursor
- Ollama installed and running
- At least one chat / coder model pulled in Ollama
- Optional: an embedding model (for example
nomic-embed-text) for richer workspace retrieval
Installation
- Install from the Visual Studio Marketplace Pre-Release channel (or install the official VSIX when provided).
- Install and start Ollama.
- Pull a recommended model (see below).
- Open the Reflexor Chat view or start typing to use inline completion.
First run
On first activation Reflexor:
- starts the local backend
- checks Ollama availability
- may recommend models based on your hardware / policy settings
If the backend or Ollama is offline, use Output → Reflexor and the status bar entry for recovery hints.
Recommended models
Typical alpha recommendations (exact suggestions can vary by hardware policy):
| Role |
Example model |
| Chat / Agent |
qwen2.5-coder:3b (or policy-selected alternative) |
| Inline completion (qualified 4 GB policy) |
qwen2.5-coder:1.5b |
| Embeddings / RAG |
nomic-embed-text |
Models are licensed by their providers, not by Reflexor.
Hardware expectations
- Reflexor is validated for this alpha on Windows x64.
- Performance depends on CPU/RAM/GPU and the models you load.
- A qualified 4 GB operating policy exists for constrained machines; it is not the universal default (
off unless you enable qualification).
- Linux and macOS are not release-validated in this alpha.
Inline completion
Ghost completion uses local Ollama models. Under a qualified hardware policy, Reflexor selects the policy Ghost model (commonly qwen2.5-coder:1.5b) unless you set an explicit manual override. You can inspect effective Ghost settings from the Reflexor debug/status commands when debug tooling is enabled.
Chat
Use Ask for explanations, Plan for written plans, and Agent for experimental change workflows. Conversation history and prompts stay on your machine for local Ollama turns.
Workspace context
Reflexor can index your workspace for retrieval-augmented answers. Indexing uses local embeddings and local storage. Large workspaces may take time; indexing yields to interactive completion when you are actively typing.
Experimental Agent
Agent mode can propose multi-file edits and commands. Treat it as experimental:
- review diffs before accepting
- approve writes / runs unless you deliberately enable broader auto-approve
- expect limitations and occasional incorrect plans
Web Search and Web Research
Core model execution supports local Ollama.
Web Search and Web Research are enabled by default and may make external requests when those features are used.
- Web Search may send your search query text to the configured provider.
- URL Fetch / Web Research uses an explicit ask permission mode before fetching pages (unless you change that setting).
- Telemetry is disabled by default.
Turn off reflexor.webSearch.enabled and reflexor.webResearchEnabled in Settings for offline-only sessions after Ollama and models are installed.
Reflexor does not claim “100% offline”, “no data ever leaves the machine”, or “fully private under every configuration”.
Privacy
See PRIVACY.md for the full policy.
Summary:
- Local Ollama + local backend for core chat / completion / indexing
- Web features may leave the machine when used
- Telemetry off by default
- MCP off by default
Troubleshooting
| Symptom |
What to try |
| Backend offline |
Reload the window; check Output → Reflexor; free local ports used by Reflexor |
| Ollama offline |
Start Ollama; wait for the status bar to recover |
| No models |
Pull a coder model; use Reflexor model install / recommend flows |
| Want offline-only |
Disable Web Search and Web Research in Settings |
| Slow completion on 4 GB class machines |
Enable / follow the qualified policy; prefer the 1.5b Ghost model; avoid oversized context overrides |
| Indexing storms |
Pause heavy indexing while typing; ensure embedding model is installed |
Known limitations
- Windows x64 only for this alpha validation
- Linux and macOS not release-validated
- Agent workflows are experimental
- Ollama and models must be installed separately
- Performance varies widely by hardware and model choice
- Interactive Marketplace operator sign-off may still be required before public pre-release upload
Support
License
Reflexor is proprietary software. It is free to use in this version under LICENSE.txt, together with TERMS.md and PRIVACY.md.
Third-party components remain under their own licenses — see THIRD_PARTY_NOTICES.txt.
Source code is maintained privately and is not open source.
