The official Harness MCP server that routes 11 streamlined tools across 139 resource types, giving AI agents programmatic access to the full Harness.io platform including CI/CD, GitOps, Feature Flags, Cloud Cost Management, Security Testing, and more.
Harness MCP Server is the official Model Context Protocol (MCP) server implementation from Harness Inc., currently at version 0.9.0, written in TypeScript and running on Node.js ≥20.
The core design philosophy is "streamlined toolset + registry-based routing": only 11 semantically clear MCP tools (list/get/create/update/delete/execute/search/diagnose/schema/status/describe) are exposed to the LLM, routed internally via a registry to 139 Harness resource types. This effectively avoids tool selection degradation caused by tool proliferation. Adding new resource types requires only declarative data files without modifying tool registration logic or schemas.
Platform coverage spans 30 toolsets: CI/CD Pipeline management, GitOps (Argo CD), Feature Flags (including Split.io FME), Cloud Cost Management (CCM), Security Testing Orchestration (STO), Chaos Engineering, Internal Developer Portal (IDP), Software Supply Chain Assurance (SCS), Software Engineering Insights (SEI with DORA metrics), Governance, Access Control, and more.
The transport layer supports stdio (for local clients like Claude Desktop / Cursor / Windsurf) and HTTP (session-based + SSE for remote/shared deployments and various MCP Gateways). Security mechanisms include write-operation Elicitation user confirmation, read-only mode, structured audit log output to stderr, client-side rate limiting, and automatic retries. Deployment options include zero-install npx startup, Docker containers, and Kubernetes (with complete manifests including non-root security context). 27 built-in prompt templates cover high-frequency scenarios such as build/deploy, troubleshooting, cost optimization, and vulnerability classification. Zero-configuration startup requires only HARNESS_API_KEY, with Account ID auto-extracted from PAT tokens, supporting dynamic multi-project discovery and automatic Harness UI URL parsing.
Installation#
npx zero-install (recommended)
HARNESS_API_KEY=pat.xxx.xxx.xxx npx harness-mcp-v2@latest
HARNESS_API_KEY=pat.xxx.xxx.xxx npx harness-mcp-v2@latest http --port 8080
Global install
npm install -g harness-mcp-v2
harness-mcp-v2
Docker
docker run --rm -p 3000:3000 -e HARNESS_API_KEY=pat.xxx.xxx.xxx harness-mcp-server
Kubernetes
kubectl apply -f k8s/
Key Environment Variables#
| Variable | Required | Default | Description |
|---|---|---|---|
HARNESS_API_KEY | ✅ | — | Harness PAT or Service Account Token |
HARNESS_ACCOUNT_ID | No | Extracted from PAT | Only needed for non-PAT keys |
HARNESS_BASE_URL | No | https://app.harness.io | Override for self-hosted Harness |
HARNESS_ORG | No | default | Default organization ID |
HARNESS_PROJECT | No | — | Default project ID |
HARNESS_TOOLSETS | No | All | Comma-separated toolset list |
HARNESS_READ_ONLY | No | false | Read-only mode |
HARNESS_SKIP_ELICITATION | No | false | Skip write-operation confirmation |
HARNESS_API_TIMEOUT_MS | No | 30000 | API timeout |
HARNESS_MAX_RETRIES | No | 3 | Retry count (429/5xx) |
HARNESS_RATE_LIMIT_RPS | No | 10 | Client-side rate limit |
Unconfirmed Information#
- npm package
harness-mcp-v2publish status not directly verified in the repo - Complete list of 139 resource types pending confirmation from
data/schemas/directory - v1.0 GA timeline not yet defined
- Split.io (FME) specific permission configuration details not expanded
- Whether HTTP transport has additional request-level authentication beyond environment variables is unclear