agentcore-dingtalk-mcp
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
agentcore-dingtalk-mcp
Talk to DingTalk as yourself — from Amazon Quick and any other MCP host. Runs on Amazon Bedrock AgentCore in AWS Global and AWS China regions.
Based on keithyt06/quick-dingtalk-mcp (MIT). This repository adds an AWS China deployment flavour (API Gateway direct, no CloudFront), keeps the AWS Global flavour (CloudFront, optional WAF), and adds a CodeBuild image pipeline and per-user AgentCore sessions. Internal resource names (
QdmRemote*stacks,/qdm-remote/*parameters,quick-dingtalk-mcp/users/*secrets) are unchanged so existing deployments upgrade in place.
An MCP server that wraps DingTalk's official CLI (dws) so an AI assistant can send and read DingTalk messages with your real user identity — your avatar, your name in the group — instead of posting as a bot.
English · 中文
Why this exists
DingTalk's own MCP server (open-dingtalk/dingtalk-mcp) only speaks as a bot — your teammates see a robot talking in the group, not you. That's wrong for a personal assistant.
This project takes the opposite route. It wraps dws, DingTalk's official CLI, which authenticates over user-identity OAuth. So when you tell your assistant "post 'meeting moved to 3pm' in the project group", the message lands authored by you — exactly as if you'd typed it.
dws is itself a thin client to DingTalk's MCP gateway (mcp-gw.dingtalk.com), meaning DingTalk's backend is already MCP-native. This project exposes that capability in two ways:
Local MCP | Remote MCP | |
Who it's for | Just you, on your own machine | A whole team, shared |
Runs where |
| AWS Bedrock AgentCore + Lambda, HTTPS |
Setup |
| Admin deploys once; each teammate self-authorizes |
Identity | Your own logged-in | Per-user OAuth, isolated per teammate |
Best when | Personal use, fastest start | Many users, central audit, no local install |
Both expose the same 38 tools (see Tools) backed by the same shared command catalog.
Related MCP server: lark-cli-mcp
Local MCP
Run it on your own machine and wire it into Amazon Quick Desktop (or Claude Desktop, Cursor) over stdio.
You type Amazon Quick Desktop quick-dingtalk-mcp dws CLI DingTalk
"post in X group" ──→ (MCP host) ──stdio──→ node packages/local/server.mjs ──exec──→ dws chat ... ──HTTPS──→ mcp-gw.dingtalk.com1. Install the DingTalk CLI
npm install -g dingtalk-workspace-cliYour DingTalk org must have CLI access enabled. Admins: open-dev.dingtalk.com → "CLI 访问管理" → enable (once, org-wide). Members hitting a not-enabled wall get a one-click request prompt at login.
2. Get the project
git clone https://github.com/cn-ljh/agentcore-dingtalk-mcp.git
cd agentcore-dingtalk-mcp
npm install3. Log in to DingTalk (as you)
dws auth login --device # device flow — works over SSH / headless tooScan the code with your DingTalk app to authorize. This is the step that makes messages post as you.
4. Connect Amazon Quick Desktop
First grab two absolute paths (Quick Desktop's child process doesn't inherit your full PATH, so absolute paths are the reliable choice — and no spaces in the path):
which node # e.g. /opt/homebrew/bin/node
echo "$(pwd)/packages/local/server.mjs" # the server entrypointThen in Amazon Quick Desktop → Settings → Capabilities → MCP → + Add MCP:
Field | Value |
Connection type | Local |
ID |
|
Name |
|
Command | the |
Arguments | the |
Save. You should see quick-dingtalk-mcp · 38 tools · Connected ✅.
Once connected, asking it to look you up returns your real DingTalk identity — proof the chain works end to end:
5. Try it
In a Quick Desktop chat:
List my recent DingTalk chats, then post a markdown message titled "test"
with body "hello from quick-dingtalk-mcp" to chat_id=cidXXXX.Full walkthrough (incl. Claude Desktop / Cursor, troubleshooting, the "is it really me?" verification) → packages/local/docs/setup.md
Remote MCP
A shared, multi-user deployment on AWS Bedrock AgentCore. One deploy serves any number of teammates — onboarding a new person is self-service, with zero code changes and no redeploy.
Each teammate Amazon Quick Desktop AWS
authorizes once ──→ (MCP host, HTTPS + Bearer) ──→ [CloudFront, Global only] → API GW → mcp-middleware (Lambda)
│ verify HMAC token, load *this user's* DingTalk token
▼
AgentCore Runtime (container running dws, per-user config)
│
▼
DingTalk (authored by that teammate)Each teammate's DingTalk token is stored, KMS-encrypted, per userId in Secrets Manager; the container gives each user an isolated dws config. DingTalk tokens auto-refresh on a schedule, and your connection stays valid as long as you keep using it — the backend tracks a per-user activity window (90 days idle before re-auth is needed), so a one-time setup just keeps working.
Two deployment flavours (AWS Global and AWS China)
The Remote stack is one codebase with two deployment flavours. AWS_REGION selects the flavour:
AWS Global | AWS China ( | |
Public entry | CloudFront (caching disabled) → API Gateway HTTP API | API Gateway HTTP API, exposed directly. No CloudFront |
WAF | Optional | Not available (WAFv2 cannot attach to an HTTP API). API Gateway stage throttling instead |
Custom domain | CloudFront alternate domain, ACM cert in us-east-1 | API Gateway custom domain, ACM cert in the deploy region, ICP-filed domain |
ARNs / DNS suffix |
|
|
Image build | Local Docker or CodeBuild | CodeBuild recommended (native ARM64, base image from ECR Public) |
Guide |
OAuth, Lambdas, DynamoDB, Secrets Manager and the AgentCore Runtime are identical in both. Each user is pinned to one AgentCore runtime session (one microVM), which is reclaimed after 5 minutes idle. Override the entry point with INGRESS=apigw|cloudfront (cloudfront is rejected in China).
Admin: deploy once
git clone https://github.com/cn-ljh/agentcore-dingtalk-mcp.git && cd agentcore-dingtalk-mcp && npm install
AWS_REGION=us-east-1 bash packages/remote/scripts/deploy.sh # AWS Global: CloudFront + API Gateway
AWS_REGION=cn-northwest-1 bash packages/remote/scripts/deploy.sh # AWS China: API Gateway onlydeploy.sh provisions everything (CloudFront on Global, API Gateway, Lambdas, DynamoDB, Secrets Manager, the AgentCore Runtime container) and prints:
the authorize URL to hand to teammates
the MCP endpoint (
https://<domain>/mcp) they paste into Amazon Quick Desktop
Prereqs: a DingTalk app (AppKey/AppSecret) with
<domain>/callbackregistered as its redirect URL, AWS credentials for the target partition (AWS_REGIONis required; any AgentCore region), Docker or the CodeBuild image stack, and Node ≥ 22.6. Chicken-and-egg (you don't know the domain before deploying)? Rundeploy.sh --only-oauthfirst — it deploys just the gateway, prints the base URL (CloudFront domain on Global, API Gateway endpoint in China) to register with DingTalk, then re-run the full deploy. Updating later? Just re-rundeploy.sh— it's idempotent (the signing key and stored secrets are preserved). Full deployment guide (prereqs, two-phase first deploy, updates, smoke tests, teardown) → docs/remote-deploy.md · day-2 operations → docs/remote-operations.md.
Teammate: connect your MCP client
The gateway is a standard OAuth 2.1 Authorization Server, so your client can connect in one of two modes. They differ only in how your client obtains the credential — both reach the same backend (which runs dws under your DingTalk identity) and behave identically once connected. Pick by what your client supports:
Mode A — OAuth wizard (recommended) | Mode B — manual Bearer (fallback) | |
Use when | your client has an OAuth login (Amazon Quick does) | your client only lets you set a fixed header |
What you do | fill a few URLs, click Approve once | open a URL, copy a |
Token upkeep | client auto-refreshes — you never touch a token | paste once; stays valid while in use (re-auth after 90 days idle) |
Mode A — OAuth wizard (recommended)
In Amazon Quick: Connectors → Add connector → MCP (not Settings → MCP), choose User authentication, and fill:
Field | Value |
MCP endpoint |
|
Authorization URL |
|
Token URL |
|
Client ID |
|
Client Secret |
|
Scope |
|
Save → Quick pops a login button → approve with your DingTalk account → it wires up the token and shows 38 tools. You authorize once; Quick auto-refreshes the token from then on.
Client IDmust be exactlyquick— the gateway pre-registers this client (Quick's form requires an ID and doesn't self-register via DCR). The secret isn't checked (PKCE is). Make sure the domain is identical in all three URLs — one mistyped character breaks the login page.
Mode B — manual Bearer (fallback)
For clients without an OAuth wizard: open https://<domain>/authorize in a browser, approve with your DingTalk account, and copy the Bearer … it returns. In your client set Connection type: Remote / HTTP, transport: streamable-http, URL = https://<domain>/mcp, and header Authorization: Bearer <that value>.
Then verify (either mode): ask "use dingtalk to look up my own profile" — it returns your real org/department. You never touch the DingTalk developer console; the app is the admin's, you just authorize with your own account.
Step-by-step onboarding (recommended OAuth path, beginner-friendly) → docs/remote-新人首配-oauth.md · technical reference (transport, troubleshooting matrix, admin setup) → docs/remote-quick-desktop.md
Tools
38 tools, identical across Local and Remote, backed by a shared catalog that covers all 261 dws commands:
Group | Count | What |
Named tools | 30 | The common ones, exposed directly: |
Discover + invoke | 2 |
|
Aliases | 6 | Older names kept working, marked |
Covers IM, contacts, calendar, docs/drive, todo, DING, attendance, OA approval, AI table, minutes, mail, and more. Every DingTalk message requires a title (unlike Feishu) — the catalog enforces it.
Comparison
This project | Custom robot webhook | ||
Posts as | You (real user) | A bot | A robot |
Auth | User OAuth ( | App credentials | Webhook URL |
Read history / search | ✅ | ❌ / limited | ❌ |
Group display | Your avatar + name | Bot | Robot name |
Want the assistant to be you? Use this. Want clearly-labelled automation? Use the bot/webhook routes.
Docs
Full index (by audience): docs/README.md
Local: setup · identity verification
Remote: AWS Global deployment guide (中文) · AWS China deployment guide (中文) · first-time setup · OAuth wizard, recommended (中文) · connect help · manual (中文) · Quick Desktop onboarding · security · operations · observability · FAQ · cost
Contributing
Issues and pull requests are welcome — bug reports, new region availability notes, doc fixes, and tool-catalog improvements are all fair game. Before sending a PR:
npm run test:all # all tests must stay green (Node ≥ 22.6), incl. CDK synth tests for both Global and ChinaA few house rules: packages/shared/catalog.json is generated (npm run build:catalog), never hand-edited; changes in packages/shared/ affect both Local and Remote, so run the full suite; never commit real AppKeys, tokens, account IDs, or live domains.
Acknowledgments
keithyt06/quick-dingtalk-mcp — the upstream project this repository is based on.
dws— DingTalk-Real-AI/dingtalk-workspace-cli — does the heavy lifting; this project is a thin shim over it.
License
MIT © Keith Yu
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Hosted MCP for Google Workspace, Atlassian and Notion: one OAuth, per-user accounts, write access.
MCP server connecting AI agents to 100+ apps (Gmail, Slack, Notion, GitHub) via one-click OAuth.
Your OpenWork org's skills, plugins, workflows, and connections through one OAuth MCP URL.
Connect any AI agent to 1,000+ apps and 27,000+ actions through one remote MCP server (OAuth).
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceRemote MCP server providing 200+ Feishu tools (messages, calendar, Bitable, etc.) for Amazon Quick Desktop, hosted on AWS Bedrock AgentCore with multi-user OAuth isolation.8MIT
- AlicenseAqualityNot gradedmaintenanceEnables AI clients to send, read, reply, and search messages on Feishu/Lark as the user's own identity using MCP tools.716 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables existing apps to expose their users and data to external AI agents via MCP with OAuth 2.1 authentication and scoped tools.51 npmMIT
- AlicenseNot gradedqualityCmaintenanceEnables Amazon Quick users to access a third-party MCP/REST API through Bedrock AgentCore Gateway with per-user Entra ID 3LO authentication. A request interceptor mints and injects a per-user downstream bearer token so the model never sees it.MIT No Attribution