camunda-mcp
MCP server for Camunda 7 that lets you inspect and manage process definitions, instances, variables, tasks, incidents, history, and evaluate DMN decisions. It supports read-only operations by default with optional write capabilities, and configurable authentication.
README
@deitum/camunda-mcp
An MCP server for Camunda 7: BPMN and DMN definitions, running instances, variables, user tasks, incidents, history, DMN evaluation — and, behind a flag, the operations that change the engine.
It talks to the engine's REST API, so it works with a standalone distribution, a Spring Boot
application with camunda-bpm-spring-boot-starter-rest, and an engine published behind a reverse
proxy alike. Authentication covers the four schemes those deployments actually use: none, HTTP
Basic, a ready-made token, and an OAuth2 / OIDC token endpoint — with the token sent either as
Authorization: Bearer or in a cookie.
npx @deitum/camunda-mcp
It speaks MCP over stdio and is configured entirely through environment variables.
Setting it up in a client
Claude Desktop (claude_desktop_config.json), Cursor, Windsurf and anything else that
reads the same shape:
{
"mcpServers": {
"camunda": {
"command": "npx",
"args": ["-y", "@deitum/camunda-mcp"],
"env": {
"CAMUNDA_BASE_URL": "http://localhost:8080/engine-rest",
"CAMUNDA_USERNAME": "demo",
"CAMUNDA_PASSWORD": "demo"
}
}
}
}
Claude Code:
claude mcp add camunda \
--env CAMUNDA_BASE_URL=http://localhost:8080/engine-rest \
--env CAMUNDA_USERNAME=demo \
--env CAMUNDA_PASSWORD=demo \
-- npx -y @deitum/camunda-mcp
VS Code (.mcp.json / .vscode/mcp.json), where the password is prompted for rather than
written down:
{
"inputs": [
{
"id": "camunda-password",
"type": "promptString",
"description": "Camunda password",
"password": true
}
],
"servers": {
"camunda": {
"command": "npx",
"args": ["-y", "@deitum/camunda-mcp"],
"env": {
"CAMUNDA_BASE_URL": "http://localhost:8080/engine-rest",
"CAMUNDA_USERNAME": "demo",
"CAMUNDA_PASSWORD": "${input:camunda-password}"
}
}
}
}
A variable left as a literal ${input:NAME} counts as unset, so a placeholder that was never
filled in reads as a missing value rather than being sent to the engine as a username.
Configuration
| Variable | Required | Meaning |
|---|---|---|
CAMUNDA_BASE_URL |
yes | Engine REST root (which one?). |
CAMUNDA_AUTH |
– | none | basic | bearer | oauth. Inferred when omitted. |
CAMUNDA_USERNAME |
– | User, for Basic auth or the OAuth password grant. |
CAMUNDA_PASSWORD |
– | Its password. |
CAMUNDA_TOKEN |
– | A ready-made token to send as-is (bearer). |
CAMUNDA_TOKEN_URL |
– | OAuth2 / OIDC token endpoint (oauth). |
CAMUNDA_CLIENT_ID |
– | OAuth client. |
CAMUNDA_CLIENT_SECRET |
– | Only for a confidential client. |
CAMUNDA_GRANT_TYPE |
– | password | client_credentials. Inferred from whether a user is given. |
CAMUNDA_SCOPE |
– | Scope to request, if the provider needs one. |
CAMUNDA_AUTH_TRANSPORT |
– | header (default) | cookie — how the token reaches the engine. |
CAMUNDA_COOKIE_NAME |
– | Cookie to put it in when the transport is cookie (default JWT). |
CAMUNDA_ALLOW_WRITE |
– | true registers the tools that change the engine. Default off. |
CAMUNDA_MAX_RESULTS |
– | Default page size (default 20, hard cap 100). |
CAMUNDA_TIMEOUT_MS |
– | Per-request budget (default 30000). |
Authentication
Set only the variables your engine needs; the mode follows from them, and CAMUNDA_AUTH is there
for the rare case where the guess is wrong.
None — a local distribution with the auth filter switched off:
CAMUNDA_BASE_URL=http://localhost:8080/engine-rest
HTTP Basic — what Camunda's own ProcessEngineAuthenticationFilter speaks, and what
camunda-bpm-platform:run ships with:
CAMUNDA_USERNAME=demo
CAMUNDA_PASSWORD=demo
A ready-made token — one you already have from somewhere else:
CAMUNDA_TOKEN=eyJhbGciOi…
OAuth2 / OIDC — this server fetches tokens itself and renews them as they expire, sharing one grant across concurrent tool calls and using the refresh token when the provider issues one:
# password grant
CAMUNDA_TOKEN_URL=https://idp.example/realms/demo/protocol/openid-connect/token
CAMUNDA_CLIENT_ID=camunda-client
CAMUNDA_CLIENT_SECRET=… # confidential clients only
CAMUNDA_USERNAME=user
CAMUNDA_PASSWORD=…
# client-credentials grant — drop the user, that is the whole difference
CAMUNDA_TOKEN_URL=https://idp.example/realms/demo/protocol/openid-connect/token
CAMUNDA_CLIENT_ID=camunda-service
CAMUNDA_CLIENT_SECRET=…
Credentials are all-or-nothing per mode: a half-filled set fails at startup, naming what is missing, rather than turning every tool call into the engine's login page.
Bearer header or cookie?
By default a token travels as Authorization: Bearer <token>, which is what a Camunda engine
expects. Some deployments publish the engine behind a proxy or an OIDC filter that reads the token
from a cookie and ignores the Authorization header entirely — for those:
CAMUNDA_AUTH_TRANSPORT=cookie
CAMUNDA_COOKIE_NAME=JWT # the default; change it if the filter uses another name
When the transport is cookie no Authorization header is sent at all. The symptom of getting
this wrong is a 302 to the identity provider while the same token works in the other position —
the error message says exactly that, and which variable to flip.
Which base URL?
- standalone distribution or Spring Boot starter →
https://host/engine-rest - deployment where only the webapp is exposed →
https://host/camunda/api/engine/engine/default
Both serve the same API. Getting it wrong is the common failure and does not look like one — the
front-end that owns the wrong path tends to answer 200 text/html — so the server checks for that
and says which URLs to try.
TLS
An internally hosted engine is often signed by a private root that Node does not ship, so the first
call fails with SELF_SIGNED_CERT_IN_CHAIN while curl against the same URL works (it uses the
system store). Either point NODE_EXTRA_CA_CERTS at the root certificate, or set
NODE_TLS_REJECT_UNAUTHORIZED=0 for this process. The error message says as much rather than
repeating fetch failed.
Tools
Read-only, always registered:
| Tool | What it answers |
|---|---|
camunda_list_decision_definitions |
deployed DMN decisions (Cockpit's «Decisions») |
camunda_get_decision_dmn |
the DMN XML — the decision table itself |
camunda_evaluate_decision |
runs a decision against inputs, returns the output rows |
camunda_list_decision_instances |
evaluation history with inputs/outputs |
camunda_list_process_definitions |
deployed BPMN processes |
camunda_get_process_bpmn |
the BPMN XML (source of activity ids) |
camunda_list_process_instances |
running instances, by key or business key |
camunda_get_activity_instances |
where an instance is sitting right now |
camunda_list_variables |
an instance's variables, unwrapped |
camunda_list_tasks |
open user tasks |
camunda_list_incidents |
open incidents |
camunda_get_job_stacktrace |
the exception behind a failed job |
camunda_list_history_process_instances |
finished instances |
camunda_list_history_activities |
one instance's audit trail |
camunda_list_deployments |
deployments, newest first |
camunda_get_deployment_resource |
a deployment's files, or one file's content |
camunda_get |
escape hatch: any engine GET by path |
Registered only with CAMUNDA_ALLOW_WRITE=true: camunda_start_process_instance,
camunda_complete_task, camunda_modify_process_instance, camunda_set_variables,
camunda_send_message, camunda_set_job_retries, camunda_resolve_incident,
camunda_delete_process_instance.
The write half is not registered rather than merely discouraged: a model cannot be told "do not touch" reliably, but it cannot call a tool it was never shown.
camunda_evaluate_decision is deliberately in the read half: it changes no engine state (only a
history entry) and it is much of the reason to point this at an engine's decision tables.
Variables are passed as plain JSON ({"amount": 100}) and typed automatically; pass the engine's
own envelope ({"amount": {"value": "100", "type": "String"}}) when you need an exact type.
Results are trimmed on the way back: list tools project each row down to the fields worth reading,
maxResults is capped at 100, XML at 40 000 characters and any single result at 60 000 — an engine
page of a few thousand rows would otherwise land in the model's context whole.
Troubleshooting
| Symptom | What it means |
|---|---|
302 to the identity provider |
The token was not accepted in the position it was sent — try the other CAMUNDA_AUTH_TRANSPORT. |
returned HTML, not engine JSON |
CAMUNDA_BASE_URL points at a front-end, not the engine REST root. |
SELF_SIGNED_CERT_IN_CHAIN |
Node does not trust the chain; see TLS. |
Incomplete <mode> credentials |
A half-filled set — the message names the missing variables. |
Engine rejected the credentials (401) |
The engine understood the credentials and refused them. |
Checking a deployment by hand, which separates a bad URL from a bad token:
curl -s -u demo:demo "http://localhost:8080/engine-rest/decision-definition?maxResults=5"
A JSON array means the base URL and the credentials are both right. HTML means the base URL is
wrong. A 302 to an identity provider means the credentials were not accepted in that position.
Using it as a library
The stdio binary is the point, but the server is exported too — for a custom transport or a test harness:
import { createCamundaServer, loadConfig } from '@deitum/camunda-mcp';
const server = createCamundaServer(loadConfig(process.env));
await server.connect(myTransport);
Development
npm install
npm run verify # lint, format, types, tests, build, packaging
npm test # vitest, no network
npm run inspector # build, then drive it with the MCP inspector
Against a throwaway engine:
docker run --rm -p 8080:8080 camunda/camunda-bpm-platform:run-latest
CAMUNDA_BASE_URL=http://localhost:8080/engine-rest \
CAMUNDA_USERNAME=demo CAMUNDA_PASSWORD=demo \
npx @modelcontextprotocol/inspector node dist/cli.js
See CONTRIBUTING.md.
Licence
MIT.
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。