v0.9.0 — ros-agent rebrand

ros-agent

Full rebrand from RoboRun to ros-agent — the MCP agent layer for robots.

What changed

• PyPI package: pip install ros-agent (was pip install roborun)
• CLI commands: ros-agent and ros-agent-mcp (was roborun / roborun-mcp)
• npm package: npx ros-agent (was npx roborun)
• MCP server name: ros-agent in initialize response
• MCP resource URIs: ros-agent://server-info, ros-agent://skills, etc.
• UI branding: updated across the dashboard
• All user-facing strings, error messages, and agent system prompts updated

No breaking changes to the MCP tool surface

All 49 tools, 8 prompts, 6 resources, and the topic template work exactly the same. Just the name changed.

Install

pip install ros-agent
ros-agent

v0.8.0 — Skills, MCP Prompts/Resources, Codebase Overhaul

The big one. RoboRun is now a complete MCP-native robot platform.

Highlights

• Skills plugin system — 4 loading paths, 5 built-in skills, skill template
• Compose skill — chain tools into reusable workflows (run_sequence, save_workflow)
• Inspect skill — robot_brief (one-call robot overview), watch_topic, diff_state
• 8 MCP prompts — guided workflows: explore-robot, safety-check, quick-start, fleet-sweep, build-workflow, and more
• 6 MCP resources + 1 template — server-info, skills, ros-graph, workflows, prompts-catalog, soul. Live topic streaming via roborun://topic/{path}
• MCP logging — real-time notifications on every tool call
• 49 total tools — 30 built-in ROS + 19 skill tools
• Route decomposition — 12 focused modules, 200-line server.py
• Generic codebase — works with any ROS 2 robot, not locked to any vendor
• Safety velocity clamping — configurable limits
• Dynamic ROS context injection — agent system prompt auto-injects live graph
• Persistent cross-agent memory — facts survive across sessions
• SOUL.md — agent behavioral identity
• Thin deps — core requires only websocket-client + websockets

Install

pip install roborun
roborun

Or use as MCP server:

{ "mcpServers": { "roborun": { "command": "roborun-mcp" } } }

Full MCP surface

	Count
Tools	49
Prompts	8
Resources	6
Resource Templates	1
Skills	5

v0.7.0 — Unified ROS MCP, DDS + Rosbridge

• 30 ROS MCP tools — full introspection, pub/sub, services, actions, params, camera, depth, velocity
• Dual transport — rosbridge WebSocket + direct DDS via ros_tap + CycloneDDS
• MCP stdio transport — roborun-mcp CLI for Claude Code and any stdio MCP client
• Works with any ROS 2 robot. No ROS install needed on the RoboRun host.

v0.6.0 — 3D Scene Builder, ROS Telemetry

• 3D scene builder — depth-based point cloud reconstruction from webcam
• ROS telemetry bridge — auto-subscribes to battery, IMU, odom, joint states
• WebSocket telemetry — real-time broadcast at ws://127.0.0.1:8766

v0.5.0 — 3D Spatial Perception, Drone Support

• Telemetry dashboard — WebSocket-powered charts for battery, altitude, speed, orientation, joints
• 3D trajectory visualization — Three.js path rendering with orbit controls
• Point cloud viewer — colored depth/LiDAR point cloud
• Drone support — MuJoCo quadrotor with PID controller, 6 new tools
• Robot type system — UI auto-adapts based on robot type
• 10 new MCP tools, new blueprints (Quadrotor, G1 humanoid, Webcam Only)

v0.4.0 — MCP Server, Native ROS 2, Fast Agent

• MCP server at /mcp — any AI client (Claude Desktop, Cursor) gets robot tools
• Native ROS 2 transport — rosbridge WebSocket to any robot, no ROS install needed
• Fast SDK agent — direct Anthropic SDK streaming with sensor pre-injection
• ZK proof layer — EZKL proofs for CLIP embeddings (EU AI Act compliance path)
• npx roborun — zero-friction install

v0.3.0 — MuJoCo Simulator, Spatial Memory

• MuJoCo physics sim — full headless simulation, trained ONNX locomotion policies
• Spatial memory — CLIP-searchable, geo-indexed, multi-robot memory store
• Walking robots — Go1 and G1 walk with real neural network controllers
• Simulator integrated into Control tab with source picker

v0.2.0 — Cosmos 3, JEPA, UX Overhaul

• Cosmos 3 Nano — 16B world model via MLX 4-bit (Apple Silicon native)
• JEPA attention heatmaps — ViT activation overlay on live webcam
• CLIP zero-shot search — real-time detection highlighting
• Webcam auto-start, collapsible agent panel, model failure toasts
• Resilient webcam pipeline — model crashes don't kill the stream

v0.1.0 — Initial Release

• WASD teleop, YOLO real-time detection, CLIP text-image search
• Dataset/episode recording from webcam or robot camera
• AI agent chat with Claude tool use and streaming
• Fleet management, task scheduler, live map
• Direct MCP skill dispatch