Dapr MCP Server

dapr-mcp-server

A Model Context Protocol (MCP) server that exposes Dapr's building blocks as tools for AI agents.

---

What it is

dapr-mcp-server is a Go MCP server that fronts a Dapr sidecar and hands AI agents a curated set of tools over every Dapr building block: state, pub/sub, secrets, service invocation, actors, bindings, distributed locks, cryptography, and conversation components. Tools register dynamically from the live component set — if your deployment has no pub/sub broker, the pub/sub tools aren't exposed.

Two transports are supported: stdio for local IDE integrations (Claude Desktop, Cursor, VS Code, Claude Code) and streamable HTTP for remote or shared deployments. Every tool call is annotated with OpenTelemetry spans, metrics, and optional log export; requests can be gated by OIDC, SPIFFE, Dapr Sentry, or a hybrid of the three.

Full documentation lives at docs.dapr.io / developing-ai / mcp.

Quick start

Run the binary next to a Dapr sidecar with dapr run (install it from Install):

dapr init

# stdio, for local MCP clients that launch the server themselves
dapr run --app-id dapr-mcp-server --resources-path resources -- dapr-mcp-server

# streamable HTTP, for remote clients
dapr run --app-id dapr-mcp-server --resources-path resources -- dapr-mcp-server --http :8080

With --http, connect any MCP client to http://localhost:8080/. The container image is meant for Kubernetes, where the Dapr sidecar injector adds the sidecar through the dapr.io/enabled: "true" and dapr.io/app-id pod annotations. For the complete walk-through (components, auth, OTEL, verification), see the getting started guide.

Install

Method	Command
Container	docker pull ghcr.io/dapr/dapr-mcp-server:latest
Binary	Download from Releases and place on PATH
From source (Go 1.26.6+)	go install github.com/dapr/dapr-mcp-server/cmd/dapr-mcp-server@latest

The server is listed in the official MCP Registry as io.github.dapr/dapr-mcp-server, so MCP clients that browse the registry can install it as a container.

Configuration

The configuration guide has the full environment variable reference.

Flag	Default	Purpose
--http <addr>	unset (stdio)	Serve streamable HTTP on this address instead of stdin/stdout. Health probes are served on /livez, /readyz and /startupz.
--health-check	false	Probe /livez of a running server and exit 0 if it answers 200, 1 otherwise. Used by the container HEALTHCHECK.
--health-check-addr <host:port>	from --http, else localhost:8080	Address --health-check probes. A wildcard host such as 0.0.0.0 is probed as localhost.
--version	false	Print the version and exit.

Logs are written as JSON to stderr, so they never mix with the stdio transport on stdout. /readyz returns 503 while the Dapr sidecar is unreachable.

Settings specific to the HTTP transport:

Environment variable	Default	Purpose
DAPR_MCP_CORS_ORIGIN	unset (no CORS headers)	Origin allowed to call the server from a browser, for example https://app.example.com. Set it only when a browser-based MCP client on another origin needs access.

Documentation

Page	Purpose
Overview	What the server is, architecture, capabilities, when to use it
Getting started	Install, configure components, run, verify
Tool reference	Schemas, inputs, outputs, and safety flags for every tool
Configuration	Environment variables, flags, and transport settings
Authentication	OIDC, SPIFFE, Dapr Sentry, hybrid mode
Observability	Traces, metrics, and logs

AI coding agents contributing to this repo: see AGENTS.md. For human contributors: see CONTRIBUTING.md.

Tools at a glance

All tools, with links to their entries in the tool reference. Core tools are always registered; conditional tools register only when a matching Dapr component exists. The server follows components that Dapr hot-reloads: each time a client lists tools or calls get_components, it checks the sidecar, adds a building block's tools when its first component appears and removes them when its last one goes, and tells connected clients the tool list changed.

Category	Tool	Registration	Notes
metadata	get_components	Core	Always call first
invoke	invoke_service	Core	Service-to-service HTTP invocation
actors	invoke_actor_method	Core	Virtual-actor method call
state	save_state	state.*	Idempotent save
state	get_state	state.*	Read-only
state	delete_state	state.*	Destructive, idempotent
state	execute_transaction	state.*	Atomic batch
pubsub	publish_event	pubsub.*	Not idempotent
pubsub	publish_event_with_metadata	pubsub.*	Adds headers/TTL
bindings	invoke_output_binding	bindings.*	External-system I/O
secrets	get_secret	secretstores.*	Single secret
secrets	get_bulk_secrets	secretstores.*	High-risk; bulk fetch
conversation	converse_with_llm	conversation.*	Delegate to a downstream LLM
crypto	encrypt_data	crypto.*	RSA encrypt
crypto	decrypt_data	crypto.*	RSA decrypt
lock	acquire_lock	lock.*	Distributed mutex
lock	release_lock	lock.*	Pair with acquire_lock

Development

go build -o dapr-mcp-server ./cmd/dapr-mcp-server
go test -race ./...
golangci-lint run ./...

See AGENTS.md for the full contributor / AI-agent playbook (build commands, tool-addition pattern, testing strategy, DCO flow).

Contributing

Contributions welcome — open issues via the issue templates, sign your commits per the DCO, and follow the pull request template. See CONTRIBUTING.md for the full workflow.

Security issues: do not open a public issue — follow the Dapr security disclosure process instead.

License

Apache 2.0 — see LICENSE.