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.

At a glance: GitHub star/fork counts are unverifiable as of this audit — the repo at github.com/pulumi/mcp-server now returns a 404, and Wayback Machine snapshots show it 404’ing as far back as January 2026. The last cached figure, from the aggregator mcpserverspot.com (~188 stars, 36 forks), can’t be confirmed against the primary source and should be treated as stale. Core CLI at v3.257.0 (Aug 2026). npm downloads run ~3,690/week (up sharply from ~800/week in May 2026 — usage primarily through the remote hosted endpoint); PulseMCP currently shows ~20.9K all-time visitors, #1,211 globally. Supported clients per current docs: Cursor, Claude Code, Claude Desktop, Windsurf, and any other MCP client with OAuth support. Listed on AWS Marketplace (free, Amazon Bedrock AgentCore supported). Part of our Cloud & Infrastructure MCP category.

The Pulumi MCP server (source repo formerly at github.com/pulumi/mcp-server, which now 404s — see “What Doesn’t Work Well” below) 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. As the original announcement put it: MCP brings “real-time resource information and infrastructure management directly into the development environment.” It doesn’t just tell you what an S3 bucket looks like — it can create one.

With a remote hosted endpoint at mcp.ai.pulumi.com/mcp and a local npm package, it’s the official MCP integration from the company that pioneered infrastructure-as-code in general-purpose programming languages — though its source repository is no longer publicly browsable (see “What Doesn’t Work Well” below). Pulumi now counts 4,500+ organizations that have used Neo to delegate infrastructure work. (Pulumi’s own September 2025 launch release separately named Snowflake, Nvidia, and BMW Group among its 3,700+ overall customers at the time — that release doesn’t say which of those customers use Neo specifically, so we don’t repeat the pairing.) CEO Joe Duffy: “The gap between generating code and getting it safely into production is the defining bottleneck of the AI era."

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), cdk-migration-plan (AWS CDK to Pulumi migration), and additional CDK prompts (cdk-migration-automated, cdk-migration-manual, cdk-migration-troubleshoot) per the current docs.

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 — “no scattered credentials,” as Pulumi describes it.

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 — now with Plan Mode, CLI agent, Integration Catalog, and AGENTS.md. 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. Pulumi CEO Joe Duffy described the vision: “Things that used to take weeks can now confidently be done in minutes." Neo analyzes the request, creates an execution plan, generates code across multiple stacks, runs tests, and creates pull requests with explanations. Werner Enterprises reduced provisioning time from 3 days to 4 hours while maintaining SOC 2 compliance, and teams now ship features 75% faster. Since February 2026, Neo reads AGENTS.md files — the same open standard managed by the Agentic AI Foundation under the Linux Foundation, already adopted by Cursor, GitHub Copilot, Windsurf, and Zed across 60,000+ open source projects — so it automatically picks up project conventions without being told. Neo also supports task sharing (read-only views for any org member), and — per Pulumi’s GitHub/Slack integration announcement — you can mention @pulumi-neo directly on GitHub PRs and issues for investigation, or @Neo in Slack channels to start or continue a task, for incident response.

Neo Integration Catalog (May 20, 2026). The Integration Catalog lets admins configure once and Neo gains access to external services through remote MCP servers: Atlassian (Jira, Confluence), Datadog, Honeycomb, Linear, PagerDuty, and Supabase — plus kubectl CLI for Kubernetes cluster access. This transforms Neo from an infrastructure code agent into a cross-platform operator that can read incident history, check dashboards, and correlate observability data while managing infrastructure changes.

Recurring Neo tasks. Neo Automations can now run on a schedule (hourly, daily, weekdays, or weekly) — drift detection, dependency version updates, compliance audits — and deliver results as pull requests. Autonomous infrastructure maintenance without human triggering.

Neo Plan Mode (April 2026). Plan Mode adds a dedicated four-phase collaborative workflow: Discovery (Neo investigates existing infrastructure, code, and dependencies), Synthesis (Neo produces a detailed plan referencing specific findings), Refinement (users challenge assumptions and request alternatives through conversation), and Approval (execution begins with full discovered context). This makes upfront deliberation a first-class workflow — Neo focuses entirely on discovery until you explicitly approve. Available for all Pulumi Cloud organizations, Plan Mode works with any task mode and pairs thorough planning with whatever execution autonomy fits the situation.

pulumi neo CLI command now GA (May 18, 2026). v3.241.0 removed the PULUMI_EXPERIMENTAL gate — pulumi neo is now a stable, default-visible command. Per Pulumi’s own announcement, it’s built with bubbletea, the interactive TUI renders agent messages, tool calls, and streaming output alongside the Pulumi Console session, and Neo’s edit filesystem tool executes locally, matching the schema and response wording of the upstream mcp-claude-code tool so the agent sees consistent output whether running on Cloud or CLI. The TUI handles user approval requests for sensitive actions. 4,500+ organizations have now used Neo. Pulumi also rapidly shipped agent-friendly improvements across CLI v3.233–v3.242: JSON output for pulumi up, pulumi preview, pulumi refresh, and pulumi destroy; pulumi cloud api describe/list for CLI introspection; webhook management and scheduled actions — all structured for programmatic consumption by AI agents.

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 — requiring zero local setup. 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).

150+ providers across every major cloud. The Pulumi Registry covers AWS, Azure, Google Cloud, and Kubernetes; per Pulumi’s own count as of an April 2025 registry expansion post, the registry passed 150 providers and 7,500+ resource types (native plus Terraform-bridged) — we couldn’t find a more recent official count, so treat “150+” as a floor rather than a current figure. Multi-cloud infrastructure gets first-class support — the same MCP server handles resources from any provider.

Policy violation detection — now with ESC environments and offline analysis. 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. Beta users report automated governance reduced policy violations by 90% on average. April 2026 brought two policy enhancements: Policy Packs can now reference Pulumi ESC environments for credentials and configuration instead of hardcoded values, and the new pulumi policy analyze command (Apr 3, 2026) evaluates existing stack state against policy packs without running the Pulumi program or making provider calls — useful for auditing infrastructure that’s already deployed.

Zero-downtime migration from five IaC platforms. Neo now handles automated, zero-downtime migration to Pulumi from AWS CDK, AWS CloudFormation, Terraform, CDKTF, and Azure ARM templates — not just the convert-terraform-to-typescript prompt. This removes a significant barrier for teams evaluating Pulumi.

AWS Marketplace listing. The remote MCP server officially supports Cursor, Claude Code, Claude Desktop, Windsurf, and “any other AI assistant that supports the Model Context Protocol with OAuth” — Pulumi’s current docs don’t name Kiro or GitHub Copilot specifically, so we’ve dropped those from our client list pending re-confirmation against a primary source. The server is also listed on AWS Marketplace (free, delivered as a Docker container at version 0.1.17, with Amazon Bedrock AgentCore support confirmed on the listing), making enterprise procurement easier. The OAuth-based authentication through Pulumi Cloud eliminates per-machine credential storage.

Bun runtime support (April 2026). Pulumi now supports Bun as a fully supported runtime (Apr 8, 2026, shipped in Pulumi 3.227.0), allowing TypeScript program execution without Node.js. This matters for MCP server users because deploy-to-aws and other code generation tools can now target Bun-based Pulumi programs — one fewer dependency in the toolchain.

Agent sprawl governance positioning. Pulumi is positioning its IaC platform as the governance answer to enterprise AI agent proliferation. With the Salesforce 2026 Connectivity Benchmark estimating the average enterprise has twelve agents today (projected 67% growth over two years), the argument is that typed infrastructure graphs provide a better reasoning surface for agents than YAML or bash scripts. Neo operates on top of IaC, ESC, and policy features as guardrails, automatically respecting security rules and maintaining audit trails.

Previous provider version docs in registry (March 2026). The Pulumi Registry now serves documentation for previous provider versions — a version-selector dropdown on first-party provider pages goes back two major versions — which helps Neo and other AI agents generate accurate code for teams pinned to older provider releases.

What Doesn’t Work Well

Low community adoption, and the MCP server repository is no longer publicly accessible. The dedicated repo at github.com/pulumi/mcp-server returns a 404 as of this audit. This isn’t new: Wayback Machine snapshots show it 404’ing as far back as January 7, 2026 — before this review was first published (March 14, 2026) — so its star count, fork count, and commit history have never been independently verifiable by us against the primary source, only inferred from third-party aggregators. The last cached figure from mcpserverspot.com is ~188 stars / 36 forks / “5 months ago” last commit, but that page shows no refresh timestamp and its numbers plausibly predate the repo going inaccessible — treat them as unconfirmed, not current fact. By contrast, Terraform MCP’s repo (1,370+ stars) and Kubernetes MCP’s repo (1,950+ stars) are both live and publicly browsable. The wider Pulumi CLI, meanwhile, has kept shipping fast (ten releases in May 2026 alone, v3.232.0 → v3.242.0, and v3.257.0 as of this audit) — so this is specifically an MCP-repo visibility problem, not a sign the company has slowed down. Current npm downloads run ~3,690/week, well up from the ~800/week we measured in May 2026. PulseMCP now shows ~20.9K all-time visitors and a #1,211 global rank — figures that don’t reconcile with what a prior audit pass recorded here, which is a reminder that aggregator stats on this page should be re-verified live each pass rather than trended from memory. Community examples, troubleshooting guides, and third-party integrations remain sparser than Terraform MCP’s ecosystem.

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. Pulumi identified a “velocity trap” where AI dev tools accelerate app creation but infra teams can’t keep pace — Neo is their answer, but it requires full Pulumi Cloud adoption. 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.” As Pulumi’s own 2026 predictions blog candidly noted: “We’re going to ship code that we’ve never read before.” 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.” The npm package is still at v0.2.0 (last published Sept 2025) despite the surrounding Neo/CLI ecosystem’s rapid pace — for a project this central to Pulumi’s AI strategy, that’s a surprisingly quiet release cadence for the npm package itself.

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 1,370+ 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 150+ 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: dogukanakkaya/pulumi-mcp-server is a community alternative providing Pulumi Automation API and Cloud API access. (A second community alternative we’d previously noted here, punkpeye/mcp-server-11, is gone — its GitHub repo now returns 404, so we’ve removed the reference.) 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 May 2026 announcements mark a genuine inflection point. pulumi neo is now GA in the CLI (no longer experimental). The Neo Integration Catalog connects Neo to Atlassian, Datadog, Honeycomb, Linear, PagerDuty, and Supabase via remote MCP — and kubectl for Kubernetes. @pulumi-neo works in GitHub PRs and issues, @Neo in Slack. Recurring Neo tasks automate drift detection and compliance audits. Ten CLI releases in four weeks restructured all Pulumi commands for agent-friendly JSON output. Pulumi has also launched the first IaC provider for NVIDIA AI Cluster Runtime and added CoreWeave + Weights & Biases integrations, positioning themselves squarely in the AI infrastructure era. 4,500+ organizations have used Neo.

The weaknesses remain. Neo still requires Pulumi Cloud. The MCP server’s own GitHub repo is now inaccessible (404) rather than merely quiet, so community-adoption stats for it can no longer be independently confirmed — while the broader Neo platform keeps accelerating, the MCP interface itself is one thin, hard-to-verify layer atop a much larger platform play. deploy-to-aws still covers only AWS. Documentation is still thin for complex scenarios. Constellation Research analyst Holger Mueller put it well: “Anything that can help platform engineers to automate infrastructure management is not only welcome, but nowadays likely essential for organizations to keep operating successfully.” 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.

This review was researched and written by an AI agent (Claude Sonnet 4.6, Anthropic) based on publicly available documentation, GitHub data, and web sources. We have not tested this MCP server hands-on. Rob Nugen maintains editorial oversight. Last updated August 14, 2026.