Connect an agent to local robot evidence
OORB exposes supported local operations through a stdio MCP server. Codex and Claude Code can use it to inspect registered projects, saved robot scans, stored episodes, and model provenance. Authentication and chat permissions remain with the native agent client.
You need an OORB-approved source installation with runfp-mcp, or a server executable supplied by OORB, plus your installed agent CLI. Request access if you do not have that installation. These instructions do not install a public package or enable a hosted ChatGPT connector.
1. Prepare the server executable
For an approved source checkout, run uv sync in that checkout. On Linux or macOS, the console script is then available at:
/absolute/path/to/oorb_fp/.venv/bin/runfp-mcp
Replace this example path with the actual executable path. For another installation, use the executable OORB provided. Use an absolute path so the client can find the environment from a different working directory.
2. Add OORB to your client
For Codex:
codex mcp add oorb -- /absolute/path/to/oorb_fp/.venv/bin/runfp-mcp
codex mcp list
For Claude Code:
claude mcp add --transport stdio --scope user oorb -- /absolute/path/to/oorb_fp/.venv/bin/runfp-mcp
claude mcp list
These commands update the client's MCP configuration. If your OORB installation uses a non-default store, include its RUNFP_HOME in the server's environment configuration. The desktop or TUI Agents view can provide configuration containing the current store; merge that entry with existing client settings.
Restart the native chat after adding the server and inspect /mcp for connection status. Confirm that OORB reports the intended store and project before asking for run analysis.
Client syntax is documented by OpenAI's Codex MCP guide and Anthropic's Claude Code MCP guide, checked October 14, 2026.
3. Start with an evidence question
Try this prompt in the connected client:
Use OORB tools to list my registered projects and saved robot interfaces. Ask me which project to inspect, then compare two stored episodes and explain the available evidence and its limits.
Project selection is explicit. An agent's working directory does not automatically choose the active OORB project.
Supported operations
| Tools | Purpose |
|---|---|
oorb_status, oorb_projects_list, oorb_interfaces_list | Inspect the store, registered projects, and saved interfaces |
oorb_maps_history, oorb_maps_read, oorb_maps_compare | Inspect and compare saved hardware or package scans |
oorb_episodes_list, oorb_episodes_compare | List and compare stored robot runs |
oorb_model_show, oorb_model_history | Read model provenance |
oorb_wandb_status, oorb_wandb_projects, oorb_wandb_collections, oorb_wandb_versions, oorb_wandb_drift | Look up evidence using the linked Weights & Biases account |
oorb_project_select changes the selected project. oorb_model_set_local, oorb_model_set_wandb, and oorb_model_clear change the model record. Keep those distinct from inspection when giving your agent instructions. Weights & Biases account linking and unlinking are not exposed as MCP tools.
Understand the current boundaries
Saved scans describe historical observations; they do not establish a robot's current state. The MCP server exposes the operations listed above. It does not provide an autonomous recording pipeline, a persistent job coordinator, or a closed-loop robot controller.
The desktop and TUI Agents views can also launch the installed native Codex or Claude Code terminal client with OORB configuration. Chats run in that client; OORB does not embed or operate their transcripts.
Read the run comparison guide before interpreting an agent's findings.