AGILEX Docs

AGILEX Docs MCP

Official product documentation for AgileX Robotics (松灵机器人) — user manuals, CAN 2.0B protocol references, firmware notes, SDK tech posts and after-sales FAQ for piper, pika, nero, ranger, limo, scout, bunker, hunter, tracer, umr, cobot_magic and aloha.

Point your AI coding assistant at this server before writing integration code for AgileX hardware, so it consults the real manual instead of guessing baud rates, frame formats or pinouts.

This repository documents the public interface. It covers the endpoint,

the tool signatures and every client configuration. The indexing service that

backs it is operated by AgileX and is not part of this repository.

---

Endpoint

No API key, no OAuth, no local process. If your client speaks Streamable HTTP, you only need the URL.

---

Quick start

Clients are listed in order of how common they are. Pick one.

# 1) Claude Code — all projects

claude mcp add --scope user --transport http agilex-docs https://service.agilex.ai/agilex-docs/mcp

2) Codex CLI — writes ~/.codex/config.toml

codex mcp add agilex-docs --url https://service.agilex.ai/agilex-docs/mcp

Cursor one-click:

cursor://anysphere.cursor-deeplink/mcp/install?name=agilex-docs&config=eyJ1cmwiOiJodHRwczovL3NlcnZpY2UuYWdpbGV4LmFpL2FnaWxleC1kb2NzL21jcCJ9

Any client that reads mcpServers (Cursor, Claude Desktop, Cline, Windsurf, Cherry Studio, Continue, Zed, LibreChat, Dify …):

{

"mcpServers": { "agilex-docs": { "url": "https://service.agilex.ai/agilex-docs/mcp" } } }

Full per-client instructions: docs/CLIENTS.md.

---

Tools

Five read-only tools. The names are deliberately verb_noun so a model can pick the right one without reading the schema first.

search_docs returns doc_id, title, space, url, score, chars and a snippet. Then use find_in_doc to jump to the passage you need and read it with get_doc(doc_id, offset=…). If the response has truncated: true, call get_doc again with next_offset.

Manuals open with a cover page and preface, so reading from offset 0 rarely answers a specific question — that is why the search-then-locate-then-read path is the recommended one rather than get_doc alone.

max_chars is honoured as requested (1–120 000), and the applied value is echoed back as max_chars in the response. There used to be an undocumented 2000-character floor; a caller paging in 1000-character steps would then re-read overlapping text without any indication, so it was removed.

Field semantics

count means different things in different tools, deliberately:

find_in_doc(limit=1) can therefore return count: 24 with one hit — returned tells you what is in the array. matched_terms / missing_terms make partial matches explicit instead of silently ignoring words the page does not contain.

Numeric parameters declare their bounds in the schema: limit 1–25 (search), 1–10 (find_in_doc), 1–300 (list_docs); max_chars 1–120000; offset ≥ 0. An out-of-range value is a validation error, not a silent clamp.

A result whose title carries a 【未发】 / 【暂不对外】 marker is still searchable — whether it is published is the documentation owner's call, and the code does not blacklist by keyword — but the entry carries status_note so a model does not cite unreleased material as settled.

Why find_in_doc exists

13% of the corpus needs pagination and the longest page is 162 000 characters; locating one CAN frame table in that by paging with get_doc costs nine round trips. find_in_doc scores paragraphs and returns their absolute offsets, so a single get_doc(doc_id, offset=…) lands on the table.

Pass the *concept* as query, not the product name. Terms that already appear in the page title are ignored: search_docs picked the page by them, so repeating them only dilutes the ranking. This is measurable — on the RANGER MINI 3.0 manual, query="RANGER MINI 3.0 CAN 波特率" returned 44 passages with "1.1 产品列表" on top, while query="CAN 波特率" returned 2 with "3.2 CAN接口协议" on top.

Scoring is weighted by corpus-wide document frequency, not raw term count: can appears in 138 pages (weight 1.25) while 波特率 appears in 39 (weight 2.29). If the query contains a distinctive term that the page does not contain, the tool returns zero hits and says so — otherwise English manuals produced 76 meaningless can matches for the query "CAN 波特率". A query made only of generic terms (e.g. just CAN) still returns matches.

Response shape

// search_docs("piper CAN 波特率")

{ "query": "piper CAN 波特率", "count": 3, "results": [ { "doc_id": "01a06659-2d4e-7029-bc64-63f9352b34a6", "title": "PIPER X 用户手册(CN)", "space": "AGX-机械臂产品手册", "url": "https://kb.agilex.ai/s/agx-robot-arm-products-cn-en/p/pWVjLJsYc0", "score": 18.42, "chars": 13264, "snippet": "...CAN通信标准采用的是CAN2.0B标准，通讯波特率为1000K..." } ] }

// get_doc(doc_id="01a06659-…", max_chars=2000)

{ "doc_id": "01a06659-2d4e-7029-bc64-63f9352b34a6", "title": "PIPER X 用户手册(CN)", "url": "https://kb.agilex.ai/s/agx-robot-arm-products-cn-en/p/pWVjLJsYc0", "chars": 13264, "offset": 0, "truncated": true, "next_offset": 2000, "content": "..." }

// find_in_doc(doc_id="01a06659-…", query="波特率")

{ "doc_id": "01a06659-2d4e-7029-bc64-63f9352b34a6", "title": "PIPER X 用户手册(CN)", "url": "https://kb.agilex.ai/s/agx-robot-arm-products-cn-en/p/pWVjLJsYc0", "chars": 13264, "query": "波特率", "count": 1, "hits": [ { "offset": 12101, // ← feed straight into get_doc "heading": "6 机械臂CAN协议", "score": 45.85, "matched": ["波特", "特率"], "excerpt": "…通讯波特率为1000K，报文格式采用MOTOROLA(MSB)格式…" } ], "hint": "用 get_doc(doc_id, offset=hit.offset) 读这一段前后；offset 是全文绝对位置。" }

url always points at the original page — cite it so the user can verify.

---

Client configuration

Claude Code

# all projects

claude mcp add --scope user --transport http agilex-docs https://service.agilex.ai/agilex-docs/mcp

this project only (writes .mcp.json, can be committed)

claude mcp add --scope project --transport http agilex-docs https://service.agilex.ai/agilex-docs/mcp

claude mcp list

</details>

Codex CLI

codex mcp add agilex-docs --url https://service.agilex.ai/agilex-docs/mcp

or ~/.codex/config.toml:

[mcp_servers.agilex-docs]

url = "https://service.agilex.ai/agilex-docs/mcp"

Inside Codex, /mcp lists active servers. Codex uses the server's instructions field as global guidance, so the search-then-read workflow is already in place.

Cursor

.cursor/mcp.json in the project, or ~/.cursor/mcp.json globally:

{

"mcpServers": { "agilex-docs": { "url": "https://service.agilex.ai/agilex-docs/mcp" } } }

Reload the window (Ctrl+Shift+P → Reload Window). Settings → MCP should list agilex-docs with 5 tools.

VS Code / Copilot

.vscode/mcp.json — note the top-level key here is servers, not mcpServers:

{

"servers": { "agilex-docs": { "type": "http", "url": "https://service.agilex.ai/agilex-docs/mcp" } } }

Or Ctrl+Shift+P → MCP: Add Server → HTTP → paste the URL.

Claude Desktop (stdio bridge)

Claude Desktop only speaks stdio, so bridge the endpoint with mcp-remote:

{

"mcpServers": { "agilex-docs": { "command": "npx", "args": ["-y", "mcp-remote", "https://service.agilex.ai/agilex-docs/mcp"] } } }

Quit and reopen the app completely after saving.

Cherry Studio

Settings → MCP Servers → Add → type streamableHttp, URL as above. Cherry Studio also accepts the mcpServers JSON directly.

DeepSeek Harness (DSH)

DSH connects through @deepseek-ai/dsh-mcp-client, configured as a Cordis overlay:

- insert:

• id: agilex-docs name: '@deepseek-ai/dsh-mcp-client' config: serverName: agilex-docs transport: streamable-http url: https://service.agilex.ai/agilex-docs/mcp

dsh web --patch "$PWD/examples/agilex-docs.cordis.yml"

Tools appear as mcp__agilex-docs__search_docs.

ChatGPT custom connector

Paid plan, web only: Settings → Connectors → Advanced → enable Developer mode → Create → URL above, authentication No authentication.

---

Example prompts

I need a CAN control script for the piper arm. First call search_docs with

"piper CAN 报文", then get_doc on the best hit, and write the code using the

frame format and baud rate from the manual. Cite the manual URL in a comment.

Our ranger mini throws a chassis error code. Search the after-sales FAQ for the

code, quote the troubleshooting steps and link the page.

Because the server ships an instructions block telling the model to search first and cite URLs, the workflow usually happens without any extra prompting.

---

Scope and data

The corpus covers product documentation that is publicly accessible at https://kb.agilex.ai. Internal spaces are excluded: the s