Openeye
Run AR/XR procedure verification sessions from an agent with memory, visuals, and optional DPO trajectory export for training pipelines.
Overview
OpenEye is a MCP server for the Build phase that verifies AR/XR procedures with visual sessions, memory, and optional DPO trajectory export.
What is this MCP server?
- Procedure verification for AR/XR with memory and visual session support
- Python sidecar entry sidecar/mcp_server.py via @dumbspacecookie/openeye npm wrapper (v0.1.1)
- Default agent path uses Anthropic (ANTHROPIC_API_KEY required); swappable providers via src/models.ts
- Optional Context training opt-in (OPENEYE_CONTEXT_OPTIN, OPENEYE_CONTEXT_API_KEY) with default off
- Package and server version 0.1.1
- Transport: stdio via npm identifier @dumbspacecookie/openeye
- OPENEYE_CONTEXT_OPTIN defaults to false
Community signal: 2 GitHub stars.
What problem does it solve?
You are building spatial or procedural copilots but lack an MCP-native loop for visual step verification and trajectory capture.
Who is it for?
Solo builders shipping AR/XR or checklist-heavy field apps who want MCP-linked multimodal agents and eval exports.
Skip if: Simple CRUD web apps with no procedural or visual verification requirements.
What do I get? / Deliverables
After setup, your agent can drive OpenEye verification sessions and optionally export trajectories when you opt into Context training.
- MCP tools for visual procedure verification sessions
- Session memory across verification steps
- Optional trajectory export to Context when opted in
Recommended MCP Servers
Journey fit
OpenEye extends how agents supervise step-by-step physical or XR procedures—primary Build placement under agent-tooling rather than generic app frontend work. Agent-tooling is the shelf for MCP servers that wrap multimodal agents, sidecars, and trajectory export for ML workflows.
How it compares
Multimodal agent MCP sidecar, not a generic browser automation or unit-test runner.
Common Questions / FAQ
Who is OpenEye for?
Developers building AR/XR or procedure-following agent experiences who use Claude Code or other MCP clients with Python sidecars.
When should I use OpenEye?
During Build agent-tooling when you need verified step sequences, session memory, and optional DPO trajectory export for research or fine-tuning.
How do I add OpenEye to my agent?
Install @dumbspacecookie/openeye, set ANTHROPIC_API_KEY, configure stdio to run python3 on sidecar/mcp_server.py, and only set OPENEYE_CONTEXT_OPTIN if you accept Context data sharing.