UniRoboSim MCP
MCP server exposing UniRoboSim evidence, simulation state, camera images, and optional explicit control of owned simulation sessions.
README
UniRoboSim MCP
unirobosim-mcp exposes UniRoboSim evidence, simulation state, backend camera images, and explicitly enabled simulation control to MCP-compatible clients. The server has two deployment profiles:
- Evidence profile (default): bounded, read-only access to an operator-selected evidence root.
- Control profile (explicit): Evidence tools plus Read and Control tools for simulator sessions created and owned by this server.
The server does not attach to sessions created by other applications.
Installation
Python >=3.11,<3.13 is supported. Install Core, this package, and the adapter required by the selected backend in the same environment.
conda create -n unirobosim-mcp python=3.12 pip -y
conda activate unirobosim-mcp
git clone https://github.com/GitHofee/UniRoboSim.git
git clone https://github.com/GitHofee/UniRoboSim-mcp.git
git clone https://github.com/GitHofee/UniRoboSim-mujoco.git # example backend
python -m pip install ./UniRoboSim ./UniRoboSim-mcp ./UniRoboSim-mujoco
General deployments use the current MCP 2.x runtime. Isaac Lab 3.0 environments retain their verified Pydantic and Uvicorn pins, so install the compatibility extra there:
python -m pip install './UniRoboSim-mcp[isaaclab]'
The extra selects MCP 1.10.1; it exposes the same UniRoboSim tool catalog and was
verified through the real stdio protocol with Isaac Sim 6.0.1.
Evidence profile
unirobosim-mcp --root /absolute/path/to/approved/evidence
UNIROBOSIM_EVIDENCE_ROOT may be used instead of --root:
export UNIROBOSIM_EVIDENCE_ROOT=/absolute/path/to/approved/evidence
unirobosim-mcp
| Tool | Contract |
|---|---|
evidence_server_info |
Return the active root, hard query limits, and control status. |
list_debug_evidence |
List allowlisted evidence with a bounded POSIX glob. |
read_debug_evidence |
Read one bounded UTF-8 or JSON artifact. |
summarize_debug_trace |
Validate a closed trace and return its compact manifest. |
query_debug_events |
Query publish, clear, and reset events without full geometry. |
query_debug_reports |
Query accepted, filtered, and dropped publish decisions. |
query_debug_primitives |
Reconstruct selected active debug primitives at a sequence. |
Absolute paths, traversal, escaping symlinks, unapproved extensions, oversized files, excessive scans, and excessive result counts are rejected.
Control profile
Control must be enabled explicitly. Local asset files are denied unless their parent tree is allowlisted with --asset-root.
unirobosim-mcp \
--root /absolute/path/to/approved/evidence \
--enable-control \
--asset-root /absolute/path/to/approved/assets \
--max-sessions 2 \
--lease-timeout-seconds 300
Read API
Read tools require a session ID but not the write lease.
| Tool | Contract |
|---|---|
simulation_list_backends |
Discover and probe installed backend entry points. |
simulation_list_sessions |
List sessions owned by this server; lease values are never returned. |
simulation_scene_snapshot |
Return the portable scene graph for entity and camera discovery. |
simulation_get_entity |
Read typed state for a rigid body, articulation, deformable, particle fluid, or camera. |
simulation_capture_camera |
Return an MCP image containing PNG data encoded from the backend RGB camera buffer. |
simulation_get_entity reports the canonical path, entity kind, original MCP configuration, simulation tick, array shapes and dtypes, and type-specific data. include_values=true includes bounded values; include_contact=true adds rigid-body contact state.
simulation_capture_camera is not a desktop or browser screenshot. It calls the selected backend through Camera.read("rgb"), validates the canonical [environment,height,width,3] uint8 buffer, and encodes that buffer as PNG. save_to_evidence=true also writes the image under <root>/screenshots/ and returns its SHA-256 digest and dimensions.
Control API
All mutations require the opaque lease_id returned by simulation_create and a unique command_id.
| Tool | Contract |
|---|---|
simulation_control_info |
Return ownership policy, allowlisted roots, and hard resource limits. |
simulation_create |
Create an owned EasyAPI session for an explicit backend. |
simulation_configure_entity |
Add a box, rigid asset, articulation, camera, deformable, or particle fluid before start. |
simulation_start |
Compile the scene and return its backend build fingerprint. |
simulation_renew_lease |
Extend the write lease without changing its value. |
simulation_step |
Advance the simulation by a bounded number of steps. |
simulation_reset |
Reset all or selected environments. |
simulation_command |
Apply articulation, rigid-wrench, deformable, fluid, scene, or debug-clear commands. |
simulation_close |
Close the owned session and release backend resources. |
Repeated use of a command_id with identical input returns the cached result with idempotent_replay=true. Reusing that identifier with different input is rejected. Expired sessions are closed automatically. Every applied or rejected mutation is written to mcp-control-audit.jsonl; lease values are excluded from the audit record.
Agent operating rule
An agent using the Control profile must follow this sequence:
- Call
simulation_list_backendsand select an available backend explicitly. - Call
simulation_create; retain the returned lease only for write operations. - Add all entities with unique command identifiers, then call
simulation_start. - Use
simulation_scene_snapshotto discover canonical entity and camera paths. - Use
simulation_get_entityfor targeted state andsimulation_capture_camerafor visual verification. - Reuse a command identifier only to retry the identical write request.
- Call
simulation_closefor every created session, including failed workflows.
The agent must not infer backend support from tool availability. Unsupported simulator capabilities are reported by capability negotiation or by the selected adapter.
Loopback HTTP
unirobosim-mcp \
--root /absolute/path/to/approved/evidence \
--transport streamable-http \
--host 127.0.0.1 \
--port 8766
Unauthenticated HTTP is restricted to 127.0.0.1, localhost, or ::1. Remote deployment requires an authenticated and authorized gateway. Control mode must not be exposed directly on an untrusted network.
Programmatic embedding
from pathlib import Path
from unirobosim_mcp import ControlLimits, EvidenceLimits, SimulationControl, create_server
root = Path("/approved/evidence")
control = SimulationControl(
root,
asset_roots=(Path("/approved/assets"),),
limits=ControlLimits(max_sessions=1, lease_timeout_seconds=120),
)
server = create_server(
root,
limits=EvidenceLimits(max_results=50, max_query_items=100),
control=control,
)
server.run(transport="stdio")
Verification
python -m pip install -e '.[dev]'
ruff format --check src tests
ruff check src tests
mypy src
coverage run -m pytest
coverage report
Release acceptance calls every published MCP tool through a real in-process MCP client. Additional contract tests cover all supported entity types and command families, leases, idempotency, expiration, allowlisted assets, resource limits, audit records, PNG encoding, and saved screenshot evidence. Native acceptance is executed separately for each installed simulator adapter; a feature is not reported as passed for a backend unless that native run succeeds.
Core contracts and adapter installation are documented in UniRoboSim Core.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。