flexorch-mcpofficial
A
licenseA
qualityB
maintenanceEnables Claude and other MCP-compatible agents to process documents, extract structured data, detect PII, and export LLM-ready datasets through natural language tool calls.
8
21 PyPI
1
MIT
embodied-mcp
by DDDB479
This MCP server provides embodied-AI dataset tooling for inspection, validation, quality analysis, annotation prep, conversion, cataloging, scaffolding, and background jobs.
Detect and summarize dataset formats such as LeRobot, HDF5/robomimic, MCAP, rosbag1/2, RLDS, and canonical JSONL.
Validate structure, metadata, features, episode consistency, fps, and format-specific rules with health scores.
Run quality reports with feature statistics, anomaly detection, caching, progress notifications, and frame/sample limits.
Evaluate configurable quality gates and plan training windows or moving/idle episode segments.
Generate annotation task manifests and, per schema, perform write operations for labeling, conversion, scaffolding, and catalog indexing.
Convert datasets to canonical JSONL or LeRobot v2.1.
Index, search, and inspect a SQLite catalog including datasets, versions, lineage, runs, history, and stats.
List supported formats, detection methods, dependencies, and available validation checks.
Submit, monitor, cancel, and list background read-only analysis jobs.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@embodied-mcp check the quality of ./datasets/aloha_pick_cube and show any issues"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Embodied Data Studio
给具身智能数据平台开发者的一套工具：格式识别 / 校验 / 质检 / 标注任务生成 / 格式转换 / 数据集编目 / 项目脚手架，同时以四种形态提供——
形态 | 入口 | 给谁用 |
Python 库 |
| 平台后端、离线脚本、CI |
命令行 |
| 终端里的快速巡检、流水线步骤 |
MCP 服务 |
| Codex、Claude Code、DeepSeek、Cursor 等任意支持 MCP 的模型/客户端 |
DSH 插件 |
| DeepSeek Harness，把上面 27 个 MCP 工具注册成原生 agent 工具 |
核心逻辑只有一份 Python 实现，模型侧无论走 MCP 还是 DSH 插件，用的都是同一套行为，不会出现"Codex 里能跑、DeepSeek 里报错"的分裂。
┌──────────────┐ ┌──────────────┐ ┌───────────────┐
│ Codex CLI │ │ Claude Code │ │ DeepSeek/DSH │
└──────┬───────┘ └──────┬───────┘ └───────┬───────┘
│ MCP (stdio) │ MCP (stdio) │ MCP │ 插件工具
└──────────┬───────┴───────────────────┘ │
▼ ▼
embodied_core.mcp_server ◄──── dsh-embodied-data (JS 桥接)
│
▼
embodied_core (Python：formats / validate / quality / convert / catalog / scaffold)30 秒上手
# 1) 装库（仓库内开发用 PYTHONPATH，正式用 pip install -e .）
pip install -e ".[all]"
# 2) 看一眼是什么格式
embodied detect ./datasets/aloha_pick_cube
# 3) 体检 + 打分
embodied validate ./datasets/aloha_pick_cube
embodied quality ./datasets/aloha_pick_cube --json
# 4) 转成 LeRobot v2.1，给训练用
embodied convert ./datasets/aloha_pick_cube ./out/lerobot --target lerobot
# 5) 把整个数据盘编目，之后按条件搜
embodied catalog index /data/embodied --db catalog.db
embodied catalog search --db catalog.db --robot so100 --min-episodes 50Python 里同样直接：
from embodied_core import open_dataset, validate
dataset = open_dataset("./datasets/aloha_pick_cube") # 自动识别 7 种格式
ids, columns = dataset.frames() # 数值列，numpy 二维
for chunk in dataset.iter_frames(chunk_frames=50_000): # 大数据集按块读取
print(len(chunk.ids), len(chunk.columns))
report = validate("./datasets/aloha_pick_cube")
print(report["score"], [i["code"] for i in report["issues"]])Related MCP server: ReleaseGuard MCP Server
能力一览
格式识别（detect_dataset）：LeRobot v2.1/v3.0、HDF5（robomimic / 通用组）、rosbag2、rosbag1、MCAP、RLDS/TFRecord、canonical JSONL。识别基于目录结构与文件魔数，不看扩展名。
校验（validate_dataset）：8 组检查——结构、元数据、特征词表、episode 一致性、fps 一致性、LeRobot 规范（含 v3 chunk/file 编号连续性）、parquet 可读性、HDF5 空 demo。每条问题带 code / severity / scope / hint，并按 error×30、warn×8、info×1 扣分，score >= 80 才建议入库。
质检（quality_report）：逐特征统计（nan/inf 比例、分位数、范数、近零比）、鲁棒异常检测（median + 1.4826·MAD，跳变/停滞/非有限值窗口）、按运动量切 moving/idle 段。记账列（episode_index、frame_index…）不参与异常检测，避免把编码器噪声当物理问题。报告按「数据集全量文件指纹 + 参数」增量缓存，数据集没变时第二次调用直接命中；帧数据按 chunk_frames 分块读取，并受 max_frames / EMBODIED_MAX_FRAMES 上限保护。
异步作业与进度（submit_job / job_status / cancel_job / list_jobs）：大质检或逐帧校验可后台运行，MCP 通过 notifications/progress 推送阶段进度，支持协作式取消。异步只开放只读分析，写入仍走同步审批路径。
标注辅助（labeling_tasks）：把异常窗口和运动段变成可交付的标注任务清单，带固定回答格式，方便回灌。
转换（convert_dataset）：任意可读格式 ⟷ canonical JSONL / LeRobot v2.1，转换时记录被丢弃的非数值特征。
编目 v2（catalog_*）：数据集、不可变版本、运行记录和版本谱系分表存储；路径作为逻辑身份，文件指纹变化时生成新版本并写 supersedes 谱系边。支持文本/格式/机器人型号/episode 数/时长/分数过滤，以及 catalog versions|lineage|runs|history 审计。
质量门（quality_gate / embodied gate）：规则可用 JSON/TOML 配置，覆盖期望 fps 与容差、必需特征 key/kind、最低分/episode/frame 数、最大 NaN 比例及异常窗口阈值。输出结构化 gate.failures，可直接作为入库或发布 CI 的退出码。
流式协议解码（decode_samples）：rosbag2 内置 CDR 读取器，覆盖 JointState、IMU、Odometry、Image/深度图、PointCloud2、LaserScan、Float/Int 数组和 String/语言指令；MCAP 安装 mcap 后复用同一解码器。RLDS 的 Example/SequenceExample 由内置 protobuf/TFRecord 轻量解码器读取，普通场景不需要 TensorFlow。
模态与媒体（inspect_media）：识别 image/video/depth/point_cloud/language，汇总任务指令；视频仅在 ffprobe 可用时探测宽高、时长、codec、帧率和码率，缺失时返回结构化安装提示。
标注闭环（labeling_tasks → apply_annotations）：任务带稳定 task_id；回写支持 valid/invalid/needs_fix、重复标签 last-wins、冲突日志、覆盖率与异常 episode 汇总，输出文件原子写入。
LeRobot v3 导出：target=lerobot-v3 生成 v3.0 chunk/file 布局、meta/stats.json 归一化统计、episode 元数据和 train/validation/test splits，并在导出后自动跑格式校验。
平台接口：只读 HTTP Catalog API 使用标准库，支持 Bearer/X-API-Key、CORS、分页和强 ETag；storage_pull 支持 file/http(s)/s3，下载带重试、SHA-256、路径穿越防护和原子发布。
v0.5 稳定性：CDR/TFRecord/点云均有消息、记录和点数上限；导出和标注写入原子化；解码器未知类型显式失败；embodied doctor 汇总依赖、解码器、资源上限和平台能力。
v0.6 生产编排：持久化 DAG 流水线（tool/processor/human/noop 节点、重试、人工等待与恢复）、人工任务队列（预标、置信度、uncertainty/stratified 抽样、分配、SLA）、处理器插件协议、版本化交付包（逐文件 SHA-256、质量门证据）和 Zarr v2 导出。
脚手架（scaffold_project）：platform（平台骨架 + 数据契约 + 质量门）/ adapter（新格式适配器 + 测试）/ pipeline（摄取流水线 + 测试），生成的文件都带类型标注，可直接跑。
文档
文档 | 内容 |
插件功能总览、工具清单、工作流与验证状态 | |
安装、第一个数据集、常见工作流 | |
支持矩阵、每种格式的探测依据、如何新增适配器 | |
27 个 MCP 工具的参数与返回结构 | |
命令、退出码、JSON 输出形态 | |
Codex / Claude Code / DeepSeek / Cursor / DSH 的接入配置 | |
DSH 插件安装、配置项、自检 | |
HTTP Catalog API、对象存储与平台诊断 | |
v0.4–v0.6 平台化、稳定性与生产编排范围 | |
插件包说明 |
目录结构
D:\具身平台插件开发\
├── src/embodied_core/ # Python 核心（唯一的事实来源）
│ ├── formats/ # 7 个格式适配器 + registry
│ ├── validate.py # 8 组检查、评分
│ ├── quality.py # 统计、异常检测、切段、标注任务
│ ├── reports.py # 质检报告 + 增量缓存编排
│ ├── gate.py # JSON/TOML 质量门规则与判定
│ ├── cache.py # 数据集指纹与 SQLite 报告缓存
│ ├── jobs.py # 后台作业、进度与取消
│ ├── ros2_codec.py # 无依赖 CDR + 常见 ROS 2 消息解码
│ ├── tfrecord.py # TFRecord / RLDS Example 轻量解码
│ ├── media.py # ffprobe 可选探测、深度/点云/语言模态
│ ├── annotations.py # 标注任务稳定 ID、回写与冲突处理
│ ├── storage.py # file/http(s)/s3 原子拉取与校验
│ ├── api.py # 只读 HTTP Catalog API
│ ├── platform.py # 依赖/能力/资源上限诊断
│ ├── pipeline.py # 持久化 DAG、人工等待、重试与统计
│ ├── task_queue.py # 人工任务、预标、抽样与分配
│ ├── processors.py # 处理器插件协议
│ ├── delivery.py # 版本化交付包与哈希校验
│ ├── convert.py # JSONL / LeRobot 目标格式
│ ├── catalog.py # schema v2：数据集/版本/运行/谱系
│ ├── scaffold.py # 项目模板
│ ├── cli.py # embodied 命令
│ └── mcp_server.py # 零依赖 MCP stdio 服务
├── packages/dsh-plugin/ # 零依赖 ESM 插件（JS 只做 MCP 桥接）
├── tests/ # unittest（pytest 兼容），181 项
├── tools/ # 夹具生成 + 端到端脚本
├── examples/ # 可直接运行的示例
└── docs/开发
python -m unittest discover -s tests -t . # 181 项单元测试，或用 pytest
python tools/e2e.py # 22 项端到端（需先 tools/make_fixtures.py 生成夹具）
node packages/dsh-plugin/scripts/probe.mjs # 插件自检（mock cordis ctx）
node --test packages/dsh-plugin/test/*.test.mjs # 17 项插件单测（路径/并发/审批）
node tools/check_clean_install.mjs --python /abs/path/to/python # 打包后临时安装并探测夹具、离线 wheel 与本地依赖树放在 .devtools/（已在 .gitignore 中）：tools/make_fixtures.py 生成 .devtools/fixtures/，pip install 不可用时把 wheel 解压到 .devtools/pydeps/ 并加进 PYTHONPATH。
Python 依赖全部可选：核心只需要 numpy；parquet 要 pyarrow，HDF5 要 h5py，MCAP 要 mcap。缺依赖时对应适配器会给出明确的 FramesUnavailable(reason, install)，而不是抛一坨 ImportError。
状态
核心库、CLI、MCP、DSH 插件均已跑通（181 项 Python 单测 + 17 项插件单测 + 22 项端到端 + 插件 probe + 干净安装）。
插件已装进本机 profile（
~/.dsh/profiles/desktop）并在 Harness 会话里实测：28 个embodied_*工具可用，live 调用 detect / summarize / quality 返回真实数据。quality_report默认写用户缓存目录；可用embodied cache stats|clear查看或清理，不占用数据集目录。iter_frames()在 JSONL、LeRobot parquet 和 HDF5 上按块产出FrameChunk；frames()仍保留，但受max_frames保护，避免误把海量数据一次性读入内存。编目库会自动把旧 v1 数据库迁移到 schema v2；重复索引未变化的文件只记一次 run，不产生新版本。
质量门规则既能内联传给 MCP，也能通过
--rules rules.json|rules.toml在 CI 中复用。npm 发布包会在
prepack时内置python/embodied_core，安装后自动解析，不再依赖仓库源码路径。写工具默认串行；工作区外目标需要一次性审批，避免转换/脚手架并发覆盖或越界写入。
rosbag2 内置 CDR 解码（JointState/IMU/Odometry/Image/PointCloud2/LaserScan/标量与语言 topic）；MCAP 在安装
mcap后可走同一解码器；RLDS 常见 Example/SequenceExample 无需 TensorFlow。视频流（LeRobot
observation.images.*）只做元数据描述，不解码像素。
Available Tools
19 toolscancel_jobARead-only
Request cooperative cancellation of a background job. Running work stops at the next analysis checkpoint; use job_status to confirm.
Inputs: job_id (required).
catalog_historyBRead-only
Inspect immutable version history, parent/child lineage and recent run records in a schema-v2 catalog, optionally filtered to one dataset path.
Inputs: db (required), path, limit.
catalog_indexA
Scan a directory tree, load every dataset found and upsert its metadata into a SQLite catalog. Incremental: unchanged datasets are skipped.
Inputs: db (required), root (required), force, validate, +1 more.
catalog_searchCRead-only
Query the dataset catalog by free text, format, robot type, episode count, duration or minimum health score.
Inputs: db (required), text, limit, format, +4 more.
catalog_statsBRead-only
Totals for schema-v2 catalogs: datasets, immutable versions, runs, lineage edges, formats, episodes, frames and duration.
Inputs: db (required).
convert_datasetBDestructive
Convert a dataset to another layout: 'jsonl' (canonical, dependency-free) or 'lerobot' (v2.1 parquet, needs pyarrow). Numeric features only; camera streams are reported, not re-encoded.
Inputs: source (required), destination (required), target, episodes, +1 more.
detect_datasetARead-only
Identify the on-disk format of an embodied-AI dataset (LeRobot, HDF5/robomimic, MCAP, rosbag1/2, RLDS, canonical JSONL) with confidence and evidence.
Inputs: path (required).
job_statusARead-only
Read a background job's status, progress, message and structured error. Set include_result=true once status is succeeded to fetch the full report.
Inputs: job_id (required), include_result.
labeling_tasksCDestructive
Generate an annotation manifest: each task names an episode, a frame range, cues and a concrete question, ready for a human reviewer or a vision-language model.
Inputs: path (required), z, out, max_tasks, +1 more.
list_formatsARead-only
List supported dataset formats, how each is detected, which optional dependency it needs, and the available validation checks.
Inputs: no named parameters.
list_jobsARead-only
List recent background jobs, newest first, optionally filtered by status (queued, running, cancelling, succeeded, failed, cancelled).
Inputs: limit, status.
plan_windowsCRead-only
Plan fixed-length training windows over the episodes of a dataset (window, stride, episode/start/end).
Inputs: path (required), limit, stride, window.
quality_gateARead-only
Evaluate configurable publish rules from an inline JSON object: expected fps/tolerance, required feature keys or kinds, minimum score/episodes/frames, maximum NaN ratio and anomaly-window counts. Returns gate failures plus validation evidence. Read-only and available as a background job.
Inputs: path (required), rules.
quality_reportBRead-only
Compute per-feature statistics (range, mean, std, percentiles, NaN ratio) and detect anomalous windows (jumps, stalls, non-finite spans) with robust z-scoring. Results are cached by dataset fingerprint and parameters unless use_cache=false. Emits MCP progress notifications when the request carries _meta.progressToken.
Inputs: path (required), z, episodes, features, +6 more.
scaffold_projectCDestructive
Write a project skeleton: 'platform' (HTTP API + quality gate + catalog + data contract), 'adapter' (new format adapter with tests) or 'pipeline' (ingest -> validate -> index).
Inputs: destination (required), kind, name, overwrite.
segment_episodesBRead-only
Split episodes into moving/idle segments using feature-speed percentiles. Useful for phase labelling and for trimming dead time.
Inputs: path (required), episode, feature, min_frames.
submit_jobARead-only
Start a long-running read-only analysis in the background. Supported tools are quality_report, validate_dataset and quality_gate. Returns a job_id immediately; poll job_status for progress and the result.
Inputs: tool (required), arguments (required).
summarize_datasetARead-only
Summarise a dataset: format, version, fps, robot type, features (key/dtype/shape/kind), episode count, frames, duration, tasks and notes. Use before any analysis to learn the schema.
Inputs: path (required), include_episodes.
validate_datasetARead-only
Run structural validation: required metadata, feature vocabulary (state/action), episode indices and lengths, fps consistency, format-specific layout rules. Returns issues with severity, stable code, hint and a 0-100 score. Emits MCP progress notifications when the request carries _meta.progressToken.
Inputs: path (required), checks, include_frames.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
B3.4/5.0
Scored across 19 tools
Disambiguation4/5
Each tool targets a reasonably distinct action (detect vs. summarize vs. validate vs. quality_report vs. quality_gate), and the descriptions draw the lines clearly (structural validation vs. statistics/anomalies vs. rule evaluation). The main soft overlaps are detect_dataset/summarize_dataset (both surface the format) and the direct-call vs. submit_job paths for validate_dataset/quality_gate, but the background job model is explained well enough to avoid misselection.
Naming Consistency4/5
Nearly all tools use a consistent snake_case verb_noun pattern (segment_episodes, detect_dataset, validate_dataset, convert_dataset, catalog_search), grouped with readable prefixes like catalog_ and job-related tools. Minor deviations are noun-phrase names such as job_status, quality_report, quality_gate and catalog_stats, which mix styles slightly but remain predictable and readable.
Tool Count4/5
At 19 tools the surface is a bit heavy, but the domain is genuinely broad: inspection, validation, quality analysis, windows/segmentation, conversion, scaffolding, cataloging and job orchestration. Almost every tool maps to a real capability rather than filler, so it lands just over the ideal band without feeling padded.
Completeness4/5
Coverage is strong end-to-end: discovery, schema summary, structural validation, quality analysis and gating, annotation manifests, conversion, project scaffolding, a searchable catalog with version history, and async job lifecycle (submit/status/cancel/list). Minor gaps remain — no tool to read individual frames/episodes or sample data, and no catalog entry deletion/update beyond upsert by index.
Maintenance
ActivityMaintained
ResponsivenessNo recent community issues
Related MCP Connectors
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Governed data discovery, exact queries, decisions, simulations, and runtime utilities over MCP.
Trusted multi-domain data for AI: 59 tools, MCP Apps, Skills, 520+ cross-domain benchmarks.
Query, join, profile, clean and convert CSV/JSON/Parquet with server-side DuckDB over MCP.
Related MCP Servers
- AlicenseAqualityBmaintenanceProvides MCP tools for scanning, redacting, and packaging dataset/model directories for public release, leveraging Presidio for PII detection and generating Hugging Face cards and EU AI Act training-data summaries.3Apache 2.0
- AlicenseAqualityBmaintenanceProvides safe, deterministic inspection, transformation, validation, and diffing of structured data (JSON, CSV, YAML, Parquet) via schema-aware MCP tools.4Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables MCP agents to discover, import, query, visualize, select, and cite NEON and local scientific datasets while keeping credentials, private paths, and bulk data out of model context.MIT