The Pulumi MCP Server — From Registry Lookups to Autonomous Infrastructure via Neo

Where the Terraform MCP server deliberately stops at documentation, Pulumi’s MCP server keeps going.

The Pulumi MCP server connects AI coding assistants to Pulumi’s full ecosystem — the registry for resource schemas and code examples, the CLI for previewing and deploying infrastructure, Pulumi Cloud for searching deployed resources across stacks, and Pulumi Neo for delegating complex multi-step infrastructure tasks to an autonomous AI agent. It doesn’t just tell you what an S3 bucket looks like — it can create one.

With 66 GitHub stars, a v1.0.0 release, and a remote hosted endpoint at mcp.ai.pulumi.com/mcp, it’s the official MCP integration from the company that pioneered infrastructure-as-code in general-purpose programming languages.

What It Does

The server operates in two modes with overlapping but distinct tool sets:

Local Mode (npm/Docker) — Registry + CLI:

Remote Mode (hosted endpoint) — Cloud + Neo:

Plus prompts for common workflows: deploy-to-aws (deploy application code), convert-terraform-to-typescript (HCL to Pulumi TypeScript), and cdk-migration-plan (AWS CDK to Pulumi migration).

Setup

Remote (recommended — no local install):

claude mcp add --transport http pulumi https://mcp.ai.pulumi.com/mcp

OAuth authentication opens a browser for Pulumi Access Token entry and organization selection. One-time setup.

Local (npm):

claude mcp add -s user pulumi -- npx @pulumi/mcp-server@latest stdio

Requires Pulumi CLI installed locally. No auth needed for registry-only usage.

Docker:

docker pull mcp/pulumi:latest

Setup difficulty: Low (remote) to Medium (local). The remote endpoint eliminates all local dependencies — just add a URL and authenticate. The local mode requires the Pulumi CLI, Node.js, and potentially cloud provider credentials. Registry lookups work without authentication in either mode.

What Works Well

Registry with code examples in real programming languages. Pulumi’s core advantage applies directly to its MCP server. When an agent looks up an aws.s3.Bucket, it gets code examples in TypeScript, Python, Go, C#, Java, and YAML — not HCL. If your team writes infrastructure in Python, the examples are in Python. The get-resource tool returns complete property schemas with input/output specifications, making type-safe code generation possible.

Resource search across all cloud infrastructure. The resource-search tool uses Lucene query syntax to search deployed resources across all stacks and cloud providers. Ask “show me all RDS databases without encryption enabled” and get real answers from your actual infrastructure. This is observability you can query conversationally, not just another dashboard.

Neo delegation for complex tasks. The neo-bridge tool is unique in the MCP ecosystem. Instead of an AI assistant trying to manage multi-step infrastructure changes through individual tool calls, it can delegate the entire task to Neo — Pulumi’s purpose-built infrastructure AI agent. Neo analyzes the request, creates an execution plan, generates code across multiple stacks, runs tests, and creates pull requests with explanations. Werner Enterprises reportedly reduced provisioning time from 3 days to 4 hours using Neo.

The full IaC lifecycle. Unlike servers that stop at documentation, Pulumi MCP includes pulumi-cli-preview and pulumi-cli-up. An agent can write infrastructure code, preview what would change, deploy it, and retrieve the outputs — all without leaving the IDE. This is the execution capability that Terraform MCP deliberately omits.

Dual local/remote architecture. The remote endpoint at mcp.ai.pulumi.com/mcp eliminates version mismatches, dependency headaches, and “works on my machine” problems. The local npm package remains available for offline work or environments where data can’t leave the network. Both modes get updates automatically (remote) or via npm (local).

170+ cloud providers. The Pulumi Registry covers AWS, Azure, Google Cloud, Kubernetes, and 170+ additional providers. Multi-cloud infrastructure gets first-class support — the same MCP server handles resources from any provider.

Policy violation detection. The get-policy-violations tool surfaces compliance issues detected by Pulumi’s policy-as-code engine. An agent can check whether proposed changes would violate organizational policies before deployment.

Terraform migration support. The convert-terraform-to-typescript prompt helps teams migrating from Terraform to Pulumi. Given HCL code, it generates equivalent Pulumi TypeScript — a practical onboarding tool for teams making the switch.

What Doesn’t Work Well

66 stars and limited community adoption. Compare this to Terraform MCP’s 1,300 stars, Kubernetes MCP’s 1,300 stars, or Docker MCP’s 900+ stars. The low star count reflects both Pulumi’s smaller market share relative to Terraform and the relatively new nature of the MCP server. Community examples, troubleshooting guides, and third-party integrations are sparse.

Neo dependency for key features. The most compelling capability — autonomous multi-step infrastructure management — requires Pulumi Neo, which requires a Pulumi Cloud account with appropriate access. Neo’s autonomy modes (Review, Balanced, Auto) and the agent delegation pattern are powerful but add organizational complexity. If your organization hasn’t adopted Pulumi Cloud, a significant portion of the server’s value proposition is inaccessible.

Pulumi ecosystem lock-in. Just as Terraform MCP only works with Terraform, Pulumi MCP only works with Pulumi. The server doesn’t help you write CloudFormation, CDK, OpenTofu, or raw Terraform. If you’re evaluating IaC tools, this server doesn’t help you compare — it’s a commitment to the Pulumi ecosystem.

AI hallucination on complex scenarios. Users report that while the MCP server and Pulumi AI work well for simple infrastructure tasks, complex or custom scenarios can produce incorrect results. The documentation acknowledges this indirectly — complex scenarios “are not well described or may not even exist.” When Neo gets it wrong on a multi-step autonomous task, debugging is harder than fixing a bad terraform plan.

SDK maturity varies by language. Pulumi supports TypeScript, Python, Go, C#, Java, and YAML, but not all provider SDKs are equally mature in every language. The MCP server inherits this limitation — code examples and schemas may be less complete for newer language bindings.

Steep learning curve for non-Pulumi users. If your team uses Terraform, the Terraform MCP server is immediately useful — it just looks up docs. Pulumi MCP assumes familiarity with Pulumi concepts: stacks, programs, state management, the Pulumi programming model. The cognitive overhead is higher.

Documentation gaps. Users report needing to study the Pulumi source code for complex SDK scenarios. The MCP server’s own documentation on third-party sites frequently shows “No documentation available.” For a v1.0.0 release, this is surprising.

deploy-to-aws is AWS-specific. Despite Pulumi’s multi-cloud strengths, the deployment prompt only targets AWS. There’s no deploy-to-azure or deploy-to-gcp. This feels like a missed opportunity and an unfinished feature set.

How It Compares

vs. Terraform MCP Server (4/5): The fundamental philosophical split. Terraform MCP is a documentation server that deliberately does not run terraform apply — safety by design. Pulumi MCP includes pulumi-cli-up and Neo delegation — execution by design. Terraform MCP has 1,300 stars and a larger user base. Pulumi MCP has code examples in real programming languages and multi-cloud resource search. Choose Terraform MCP for safe Terraform writing; choose Pulumi MCP if you want AI-driven infrastructure execution in general-purpose languages.

vs. AWS MCP Servers (4/5): AWS’s 66-server suite covers deep AWS service integration. Pulumi MCP is cloud-agnostic — one server covers AWS, Azure, GCP, and 170+ providers. AWS MCP gives you fine-grained control over specific AWS services; Pulumi MCP gives you cross-cloud infrastructure management. Complementary for AWS-heavy teams.

vs. Kubernetes MCP Server (4/5): Different layer. Kubernetes MCP manages running clusters; Pulumi MCP helps write and deploy the IaC that provisions those clusters. Pulumi MCP can deploy a Kubernetes cluster; Kubernetes MCP manages what runs on it. Use both.

vs. Docker MCP Server (3.5/5): Docker manages local containers; Pulumi manages cloud infrastructure declarations. The Pulumi MCP server is also available as a Docker image (mcp/pulumi), but the use cases are distinct.

vs. Community Pulumi MCP servers: punkpeye/mcp-server-11 provides Pulumi Automation API and Cloud API access with an experimental API. dogukanakkaya/pulumi-mcp-server is another community alternative. The official server is the one to use — it has Neo integration, the remote endpoint, and Pulumi’s backing.

The Bottom Line

Pulumi’s MCP server is the most ambitious IaC MCP server available. Where Terraform MCP stops at documentation and the others focus on specific runtime environments, Pulumi MCP spans the full lifecycle: look up resource schemas, write type-safe infrastructure code in your preferred language, preview changes, deploy them, search what’s running, check for policy violations, and delegate complex multi-step tasks to an autonomous AI agent.

The ambition is also its weakness. Neo delegation requires Pulumi Cloud adoption. The deploy-to-aws prompt covers only one cloud. Documentation is thin. Community adoption is a fraction of Terraform MCP’s. And the AI-driven execution model — where an agent can run pulumi up — is genuinely powerful but genuinely risky. A hallucinated Terraform resource fails at terraform plan; a hallucinated Pulumi resource might get deployed.

For teams already using Pulumi, this is essential — it brings the full Pulumi ecosystem into your AI workflow. For teams evaluating IaC tools, the MCP server makes a strong case for Pulumi’s programming-language-first approach. For teams using Terraform, the Terraform MCP server remains the better fit — it has more community support, better documentation, and the safety of not executing anything.

Rating: 3.5 out of 5 — the most feature-complete IaC MCP server, with unique Neo agent delegation and real execution capabilities, held back by low community adoption, Pulumi ecosystem lock-in, documentation gaps, and the inherent risk of an AI agent that can actually deploy infrastructure.