Harness MCP Server

Harness MCP Server 2.0

An MCP (Model Context Protocol) server that gives AI agents full access to the Harness.io platform through 11 consolidated tools and 218 resource types.

Related MCP server: Forage MCP Server

Why Use This MCP Server

Most MCP servers map one tool per API endpoint. For a platform as broad as Harness, that means 240+ tools — and LLMs get worse at tool selection as the count grows. Context windows fill up with schemas, and every new endpoint means new code.

This server is built differently:

• 11 tools, 218 resource types. A registry-based dispatch system routes harness_list, harness_get, harness_create, etc. to any Harness resource — pipelines, services, environments, orgs, projects, feature flags, cost data, and more. The LLM picks from 11 tools instead of hundreds.

• Full platform coverage. 38 default toolsets spanning CI/CD, GitOps, Feature Flags, Cloud Cost Management, Security Testing, Chaos Engineering, Database DevOps, Internal Developer Portal, Software Supply Chain, Infrastructure as Code Management, Governance, Service Overrides, Knowledge Graph, Visualizations, and more. Opt-in Ansible coverage is available when you need inventory and playbook data.

• Multi-project workflows out of the box. Agents discover organizations and projects dynamically — no hardcoded env vars needed. Ask "show failed executions across all projects" and the agent can navigate the full account hierarchy.

• 32 prompt templates. Pre-built prompts for common workflows: build & deploy apps end-to-end, debug failed pipelines, review DORA metrics, triage vulnerabilities, optimize cloud costs, audit access control, plan feature flag rollouts, review pull requests, approve pending pipelines, and more.

• Works everywhere. Stdio transport for local clients (Claude Desktop, Cursor, Devin Desktop), HTTP transport for remote/shared deployments, Docker and Kubernetes ready.

• Zero-config start. Just provide a Harness API key. Account ID is auto-extracted from PAT and SAT tokens, org/project defaults are optional, and toolset filtering lets you expose only what you need.

• Extensible by design. Adding a new Harness resource means adding a declarative data file — no new tool registration, no schema changes, no prompt updates.

Prerequisites

Before installing or running the server, you need a Harness API key:

1. Log in to your Harness account

2. Go to My Profile → API Keys → + New API Key

3. Create a new Token under the API key — this generates a PAT or SAT in the format <prefix>.<accountId>.<tokenId>.<secret>

4. Save the token somewhere secure — you'll need it in the next step

For detailed instructions, see the Harness API Quickstart.

Quick Start

Option 0: Hosted Harness MCP

If your Harness account has the hosted MCP service enabled, clients that support remote MCP servers can connect directly to the managed endpoint instead of running the server locally.

Important: The hosted MCP service uses Harness Platform OAuth, not HARNESS_API_KEY. It must also be enabled/configured per account by Harness Support before the endpoint can be used.

See Hosted Harness MCP for configuration examples.

Option 1: npx (Recommended)

No install required — just run it:

HARNESS_API_KEY=pat.xxx.xxx.xxx npx harness-mcp-v2@latest

Or configure the API key in your AI client (see Client Configuration below).

# Stdio transport (default — for Claude Desktop, Cursor, Devin Desktop, etc.)
HARNESS_API_KEY=pat.xxx npx harness-mcp-v2

# HTTP transport (for remote/shared deployments)
HARNESS_API_KEY=pat.xxx npx harness-mcp-v2 http --port 8080

Note: The account ID is auto-extracted from PAT and SAT tokens (pat.<accountId>... or sat.<accountId>...), so HARNESS_ACCOUNT_ID is only needed for API keys without an embedded account segment.

Option 2: Global Install

npm install -g harness-mcp-v2

# Then run directly
harness-mcp-v2

Option 3: Build from Source

For development or customization:

git clone https://github.com/harness/mcp-server.git
cd mcp-server
pnpm install
pnpm build

# Run
pnpm start              # Stdio transport
pnpm start:http         # HTTP transport
pnpm inspect            # Test with MCP Inspector

Anthropic MCP Directory bundle

The MCPB bundle manifest lives in [mcp-directory/](mcp-directory/), and the bundle icon is tracked at [icon.png](icon.png) in the repository root. Copy mcp-directory/manifest.json to the bundle root after pnpm build so the generated archive contains root-level manifest.json, icon.png, build/, package.json, and production node_modules/.

To keep the archive small, build MCPB packages from a staging directory:

pnpm prepare:mcpb

The staged package is written to dist/mcpb/ with production dependencies installed using npm's flat layout.

CLI Usage

harness-mcp-v2 [stdio|http] [--port <number>]

Options:
  --port <number>  Port for HTTP transport (default: 3000, or PORT env var)
  --help           Show help message and exit
  --version        Print version and exit

Transport defaults to stdio if not specified. Use http for remote/shared deployments.

HTTP Transport

When running in HTTP mode, the server exposes:

The HTTP transport runs in session-based mode. A new MCP session is created on initialize, the server returns an mcp-session-id header, and subsequent requests for that session must include the same header.

Operational constraints in HTTP mode:

• Set HARNESS_MCP_AUTH_TOKEN for any shared or remotely reachable deployment. When set, every POST, GET, and DELETE request to /mcp must include Authorization: Bearer <token>.

• Non-loopback binds require HARNESS_MCP_AUTH_TOKEN by default. To run unauthenticated on a non-loopback interface anyway, set HARNESS_MCP_ALLOW_UNAUTHENTICATED_HTTP=true explicitly.

• POST /mcp without mcp-session-id must be an initialize request.

• POST /mcp, GET /mcp, and DELETE /mcp for existing sessions require the mcp-session-id header.

• GET /mcp is used for SSE notifications (progress updates and elicitation prompts).

• Idle sessions are reaped after MCP_SESSION_TTL_MS milliseconds once no request or SSE stream is active (default 300000, or 5 minutes).

• GET /health is the only non-MCP endpoint.

• Request body size is capped by HARNESS_MAX_BODY_SIZE_MB (default 10 MB).

• Set x-harness-pipeline-version: 0 or 1 on the initialize request to select V0 or V1 pipeline resources for that HTTP session.

• Set x-harness-auto-approve-risk: none|low_write|medium_write|high_write|all on the initialize request to choose a stricter per-session auto-approval threshold. The server caps this value at the deployment-level HARNESS_AUTO_APPROVE_RISK, so a session can reduce but not expand the configured approval ceiling.

Multi-User Mode

Set HARNESS_MCP_MODE=multi-user for shared HTTP deployments where each client authenticates as a different Harness user. In this mode:

• HARNESS_API_KEY must not be set in the server config — the server holds no Harness credentials.

• Each session must provide x-harness-api-key on the initialize request. x-harness-account-id is required only when the API key does not embed an account segment.

• Sessions may also provide x-harness-org and x-harness-project headers to set default scope for that session.

• The Harness API key flows through to every Harness API call for that session, so the audit trail in Harness reflects the real user.

• HARNESS_MCP_AUTH_TOKEN is independent and can still be used as an additional transport-layer gate.

# Health check
curl http://localhost:3000/health

# MCP initialize request (capture mcp-session-id response header)
# In multi-user mode, x-harness-api-key is required on initialize.
# x-harness-account-id is needed only for API keys without an embedded account segment.
curl -i -X POST http://localhost:3000/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Authorization: Bearer $HARNESS_MCP_AUTH_TOKEN" \
  -H "x-harness-api-key: $HARNESS_API_KEY" \
  -H "x-harness-account-id: $HARNESS_ACCOUNT_ID" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

# Subsequent MCP request (use returned session ID)
curl -X POST http://localhost:3000/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Authorization: Bearer $HARNESS_MCP_AUTH_TOKEN" \
  -H "mcp-session-id: <session-id>" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'

# Terminate session
curl -X DELETE http://localhost:3000/mcp \
  -H "Authorization: Bearer $HARNESS_MCP_AUTH_TOKEN" \
  -H "mcp-session-id: <session-id>"

HARNESS_MCP_ALLOWED_HOSTS controls Host-header validation for DNS-rebinding protection, and CORS limits browser origins. Neither is authentication; use HARNESS_MCP_AUTH_TOKEN or an authenticated gateway/reverse proxy for access control.

Client Configuration

Note: HARNESS_ORG and HARNESS_PROJECT are optional. They set the org ID and project ID used when not specified per tool call. Agents can discover orgs and projects dynamically using harness_list(resource_type="organization") and harness_list(resource_type="project"). The deprecated names HARNESS_DEFAULT_ORG_ID and HARNESS_DEFAULT_PROJECT_ID are still accepted for backward compatibility.

Hosted Harness MCP

Harness also supports a hosted MCP endpoint for accounts that have the managed service enabled. This is useful when you want a shared remote MCP endpoint instead of running npx harness-mcp-v2 or self-hosting the HTTP transport yourself.

Important: Hosted MCP authentication uses Harness Platform OAuth. It does not use HARNESS_API_KEY in the client config. Hosted MCP availability is configured per Harness account, so you will need to work with Harness Support to enable/configure the setting before using it.

The hosted endpoint https://mcp.harness.io/mcp is a managed service. Client-side MCP config in Claude, Cursor, or Cowork cannot override which Harness environment it routes to. For Harness0 or another private Harness SaaS environment, ask Harness Support to enable/configure hosted MCP for that environment, or run the local/self-hosted server and set HARNESS_BASE_URL to the target Harness host.

Hosted MCP example:

{
  "mcpServers": {
    "harness-prod1-mcp": {
      "url": "https://mcp.harness.io/mcp",
      "auth": {
        "CLIENT_ID": "mcp-client"
      }
    }
  }
}

Example with both hosted and local entries:

{
  "mcpServers": {
    "harness-hosted": {
      "url": "https://mcp.harness.io/mcp",
      "auth": {
        "CLIENT_ID": "mcp-client"
      }
    },
    "harness-local": {
      "command": "/absolute/path/to/npx",
      "args": ["-y", "harness-mcp-v2@latest"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/directory/containing/node:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

Troubleshooting npx ENOENT or node: No such file or directory

This is a client process-launch failure, not a Harness authentication failure. The MCP server has not started yet, so changing HARNESS_API_KEY will not affect spawn npx ENOENT.

GUI apps (Cursor, Claude Desktop, Devin Desktop, VS Code) don't always inherit your shell's PATH, so they can fail to find npx or node after a config reload. Fix this by using absolute paths and explicitly setting PATH in the env block:

{
  "mcpServers": {
    "harness": {
      "command": "/absolute/path/to/npx",
      "args": ["-y", "harness-mcp-v2"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

Find your paths with which npx and which node in a terminal, then make sure the directory containing node is included in the PATH value above. Common locations:

• Homebrew (macOS): /opt/homebrew/bin/npx

• nvm: ~/.nvm/versions/node/v20.x.x/bin/npx (run nvm which current to find the exact path)

• System Node: /usr/local/bin/npx

Claude Desktop (claude_desktop_config.json)

npx (zero install)

{
  "mcpServers": {
    "harness": {
      "command": "/absolute/path/to/npx",
      "args": ["-y", "harness-mcp-v2@latest"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/directory/containing/node:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

node (local install)

npm install -g harness-mcp-v2

{
  "mcpServers": {
    "harness": {
      "command": "/absolute/path/to/harness-mcp-v2",
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/directory/containing/node:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

Claude Code (via claude mcp add)

npx (zero install)

claude mcp add harness -- npx harness-mcp-v2

node (local install)

npm install -g harness-mcp-v2
claude mcp add harness -- harness-mcp-v2

Then set HARNESS_API_KEY in your environment or .env file.

Cursor (.cursor/mcp.json)

npx (zero install, recommended for local Cursor configs)

{
  "mcpServers": {
    "harness": {
      "command": "/absolute/path/to/npx",
      "args": ["-y", "harness-mcp-v2@latest"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/directory/containing/node:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

Run which npx in a terminal and use that full path for command; include the directory from which node at the front of PATH.

node (local install)

npm install -g harness-mcp-v2

{
  "mcpServers": {
    "harness": {
      "command": "/absolute/path/to/harness-mcp-v2",
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/directory/containing/node:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

Run which harness-mcp-v2 after npm install -g harness-mcp-v2 and use that full path for command; include the directory from which node at the front of PATH.

Devin Desktop (~/.windsurf/mcp.json)

npx (zero install)

{
  "mcpServers": {
    "harness": {
      "command": "/absolute/path/to/npx",
      "args": ["-y", "harness-mcp-v2@latest"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/directory/containing/node:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

node (local install)

npm install -g harness-mcp-v2

{
  "mcpServers": {
    "harness": {
      "command": "/absolute/path/to/harness-mcp-v2",
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "PATH": "/directory/containing/node:/usr/local/bin:/usr/bin:/bin"
      }
    }
  }
}

Using a local build from source?

Replace the command with the path to your built index.js:

{
  "command": "node",
  "args": ["/absolute/path/to/harness-mcp-v2/build/index.js", "stdio"]
}

MCP Gateway

The Harness MCP server is fully compatible with MCP Gateways — reverse proxies that provide centralized authentication, governance, tool routing, and observability across multiple MCP servers. Since the server implements the standard MCP protocol with both stdio and HTTP transports, it works behind any MCP-compliant gateway with no code changes.

Why use a gateway?

• Centralized credential management — no API keys in agent configs

• Governance & audit logging for all tool calls across teams

• Single endpoint for agents instead of N connections to N MCP servers

• Access control — restrict which teams can use which tools

Docker MCP Gateway

Register the server in your Docker MCP Gateway configuration:

{
  "mcpServers": {
    "harness": {
      "command": "npx",
      "args": ["harness-mcp-v2"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx"
      }
    }
  }
}

Portkey

Add the Harness MCP server to your Portkey MCP Gateway for enterprise governance, cost tracking, and multi-LLM routing:

{
  "mcpServers": {
    "harness": {
      "command": "npx",
      "args": ["harness-mcp-v2"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx"
      }
    }
  }
}

LiteLLM

Add to your LiteLLM proxy config:

mcp_servers:
  - name: harness
    command: npx
    args:
      - harness-mcp-v2
    env:
      HARNESS_API_KEY: "pat.xxx.xxx.xxx"

Envoy AI Gateway

The server works with Envoy AI Gateway's MCP support via HTTP transport:

# Start the server in HTTP mode
HARNESS_API_KEY=pat.xxx.xxx.xxx npx harness-mcp-v2 http --port 8080

Then configure Envoy to route to http://localhost:8080/mcp as an upstream MCP backend.

Kong

Use Kong's AI MCP Proxy plugin to expose the Harness MCP server through your existing Kong gateway infrastructure.

Other Gateways

Any gateway that supports the MCP specification (Microsoft MCP Gateway, IBM ContextForge, Cloudflare Workers, etc.) can proxy this server. For stdio-based gateways, use the default transport. For HTTP-based gateways, start the server with http transport and point the gateway at the /mcp endpoint.

Docker

Build and run the server as a Docker container:

# Build the image
pnpm docker:build

# Run with your .env file
pnpm docker:run

# Or run directly with env vars
docker run --rm -p 3000:3000 \
  -e HARNESS_API_KEY=pat.xxx.xxx.xxx \
  -e HARNESS_ACCOUNT_ID=your-account-id \
  harness-mcp-server

The container runs in HTTP mode on port 3000 by default with a built-in health check.

Kubernetes

Deploy to a Kubernetes cluster using the provided manifests:

# 1. Edit the Secret with your real credentials
#    k8s/secret.yaml — replace HARNESS_API_KEY and HARNESS_ACCOUNT_ID

# 2. Apply all manifests
kubectl apply -f k8s/

# 3. Verify the deployment
kubectl -n harness-mcp get pods

# 4. Port-forward for local testing
kubectl -n harness-mcp port-forward svc/harness-mcp-server 3000:80
curl http://localhost:3000/health

The deployment runs 2 replicas with readiness/liveness probes, resource limits, and non-root security context. The Service exposes port 80 internally (targeting container port 3000).

Configuration

The server automatically loads environment variables from a .env file in the project root if one exists. Copy .env.example to .env and fill in your values. Environment variables can also be set via your shell or MCP client config.

Semantic Search

harness_search uses semantic routing to narrow scatter-gather API calls before fanning out to Harness. Three search providers are available:

Remote provider configuration:

HARNESS_SEARCH_PROVIDER=remote
HARNESS_SEARCH_SERVICE_URL=http://search-svc:8080

# Auth — any scheme via HARNESS_SEARCH_SERVICE_HEADERS (JSON object):
HARNESS_SEARCH_SERVICE_HEADERS='{"Authorization":"Bearer <token>"}'   # standard bearer
HARNESS_SEARCH_SERVICE_HEADERS='{"x-api-key":"<key>"}'               # API key header
HARNESS_SEARCH_SERVICE_HEADERS='{"x-harness-token":"<svc-token>"}'   # internal service-to-service
# Multiple headers (e.g. service mesh + tenant routing):
HARNESS_SEARCH_SERVICE_HEADERS='{"x-harness-token":"<tok>","x-tenant":"<id>"}'
# No auth (service mesh / mTLS handles it):
# omit HARNESS_SEARCH_SERVICE_HEADERS entirely

Testing the remote provider locally with the included stub service (no external dependencies):

# 1. Create a venv and install FastAPI
python3 -m venv .venv-stub
.venv-stub/bin/pip install fastapi uvicorn

# 2. Start the stub (in-memory, cosine similarity, corpus + tenant filtering)
.venv-stub/bin/uvicorn stub-search-service:app --port 8082

# 3. Build the MCP server
pnpm build

# 4. Run the integration smoke test
node test-remote-provider.mjs
# Expected output:
#   available: true
#   indexed 2 docs
#   entity search results: pipeline:ts-test score=... corpus=entities
#   knowledge search results: schema:trigger score=...
#   all-corpus search results: (merged, sorted by score)
#   isolation check (other-acct, should be empty): PASS

# 5. Tear down
kill $(lsof -ti :8082)

The stub (stub-search-service.py) implements the same /v1/health, /v1/ingest, and /v1/search contract as the production search service. It uses a simple bag-of-chars embedding so no model download is required — results are semantically plausible but not production-quality.

HTTPS Enforcement

HARNESS_BASE_URL must use HTTPS by default. If you set a non-HTTPS URL (e.g. http://localhost:8080), the server will refuse to start with:

HARNESS_BASE_URL must use HTTPS (got "http://..."). If you need HTTP for local development, set HARNESS_ALLOW_HTTP=true.

Audit Logging

All registry-dispatched Harness API operations (list, get, create, update, delete, and execute) emit structured audit events when audit sinks are configured. Mutating events include the confirmation path used by elicitation or auto-approval when a confirmation context is present; read events currently omit confirmation metadata. Local metadata and schema discovery tools that bypass the registry, such as harness_describe and harness_schema, are not part of this audit stream. A stderr sink is registered by default but goes through the normal logger and obeys LOG_LEVEL; configure file or webhook sinks for durable audit collection:

• HARNESS_AUDIT_FILE appends newline-delimited JSON events for local collection.

• HARNESS_AUDIT_WEBHOOK_URL posts { "events": [...] } batches to an HTTPS webhook, optionally with HARNESS_AUDIT_WEBHOOK_TOKEN. Failed batches are re-enqueued with bounded capacity and eventually dropped with a warning rather than blocking tool execution.

• OTEL_EXPORTER_OTLP_ENDPOINT enables audit spans when the optional OpenTelemetry peer dependencies are installed. The sink reuses an existing tracer provider when one is registered, otherwise it bootstraps a standalone OTLP exporter.

Each event includes the tool name, resource type, operation, identifiers, timestamp, risk, outcome, HTTP method/path, duration, and confirmation method when applicable. Audit sinks are best-effort telemetry; delivery issues are logged and never replay or change the underlying Harness API operation. For OTel setup details and span attributes, see specs/005-otel-audit-sink.md.

Tools Reference

The server exposes 11 MCP tools. Most API tools accept org_id and project_id as optional overrides — if omitted, they fall back to HARNESS_ORG and HARNESS_PROJECT. harness_describe is local metadata only and does not use org/project scope.

URL support: Most API-facing tools accept a url parameter — paste a Harness UI URL and the server auto-extracts org, project, resource type, resource ID, pipeline ID, and execution ID. harness_describe does not accept url.

Scope support: Resource types with account/org/project variants expose supportedScopes in harness_describe. Pass resource_scope when you need a specific level:

• resource_scope: "account" sends only accountIdentifier.

• resource_scope: "org" sends accountIdentifier and orgIdentifier.

• resource_scope: "project" sends account, org, and project identifiers.

Current multi-scope resources include connector, service, environment, infrastructure, secret, file_store, and template. If resource_scope is omitted, the registry uses the resource's default scope and configured defaults, except resources marked as optional scope may omit org/project unless explicitly passed. Harness URLs can also set the scope automatically when the path contains account-level or project-level context.

Structured output: Every tool declares an MCP outputSchema. harness_list normalizes list-like Harness responses into object-shaped structured content so strict clients can validate it: top-level arrays become { "items": [...], "total": <count>, "page": <page> }, and common wrapper keys such as content, data, body, objects, or features are hoisted to items when needed. The text response still contains the compact JSON payload returned to all clients.

Schema Lookup Workflow

Use harness_schema before creating or updating YAML-backed resources so agents can copy exact field names and constraints instead of guessing from prose.

• Bundled schemas include pipeline, template, trigger, pipeline_v1, template_v1, inputSet_v1, overlayInputSet_v1, and agent-pipeline.

• Entity schemas include connector, environment, service, secret, and infrastructure. They are scope-aware (account, org, or project) and require org_id/project_id when the selected scope requires them.

• Vendored entity snapshots are used first when they match the runtime account; otherwise the tool falls back to the Harness NG /yaml-schema API and caches the result.

• Omit path for a field/section summary, then pass a dot-separated path to inspect a nested definition.

Examples:

{ "resource_type": "pipeline", "path": "pipeline.stages" }

{
  "resource_type": "connector",
  "scope": "project",
  "org_id": "default",
  "project_id": "payments"
}

Maintainers can refresh the vendored entity snapshots with pnpm sync-entity-schemas when Harness entity YAML schemas change.

Tool Examples

Discover what resources are available:

{ "resource_type": "pipeline" }

List organizations in the account:

{ "resource_type": "organization" }

List projects in an organization:

{ "resource_type": "project", "org_id": "default" }

List pipelines in a project:

{ "resource_type": "pipeline", "search_term": "deploy", "size": 10 }

Get a specific service:

{ "resource_type": "service", "resource_id": "my-service-id" }

Run a pipeline:

{
  "resource_type": "pipeline",
  "action": "run",
  "resource_id": "my-pipeline",
  "inputs": { "tag": "v1.2.3" },
  "wait": true
}

Toggle a feature flag:

{
  "resource_type": "feature_flag",
  "action": "toggle",
  "resource_id": "new_checkout_flow",
  "enable": true,
  "environment": "production"
}

Search across all resource types:

{ "query": "payment-service" }

Diagnose an execution by ID (summary mode — default):

{ "execution_id": "abc123XYZ" }

Diagnose from a Harness URL:

{ "url": "https://app.harness.io/ng/account/.../pipelines/myPipeline/executions/abc123XYZ/pipeline" }

Diagnose connector connectivity:

{ "resource_type": "connector", "resource_id": "my_github_connector" }

Diagnose delegate health:

{ "resource_type": "delegate", "resource_id": "delegate-us-east-1" }

Diagnose a GitOps application (with options):

{
  "resource_type": "gitops_application",
  "resource_id": "checkout-app",
  "options": { "agent_id": "gitops-agent-1" }
}

Get the latest execution report for a pipeline:

{ "pipeline_id": "my-pipeline" }

Full diagnostic mode with YAML and failed step logs:

{ "execution_id": "abc123XYZ", "summary": false }

Summary mode with logs enabled (best of both):

{ "execution_id": "abc123XYZ", "include_logs": true }

Get project health status:

{ "org_id": "default", "project_id": "my-project", "limit": 5 }

List database schemas filtered by migration type:

{ "resource_type": "database_schema", "migration_type": "Liquibase" }

List database instances for a schema:

{ "resource_type": "database_instance", "dbschema_id": "my_schema" }

Get the resolved LLM authoring pipeline for a schema and instance:

{ "resource_type": "database_llm_authoring_pipeline", "resource_id": "my_schema", "dbinstance_id": "prod_db" }

List snapshot object names (e.g. tables) for a schema instance:

{
  "resource_type": "database_snapshot_object",
  "dbschema_id": "my_schema",
  "dbinstance_id": "prod_db",
  "object_type": "Table"
}

Get full snapshot metadata for specific named objects:

{
  "resource_type": "database_snapshot_object",
  "resource_id": "prod_db",
  "params": {
    "dbschema_id": "my_schema",
    "object_type": "Table",
    "object_names": ["users", "orders"]
  }
}

Pipeline Run Workflow (Recommended)

Use this sequence to reduce execution-time input errors:

1. Discover required runtime inputs

• harness_get(resource_type="runtime_input_template", resource_id="<pipeline_id>")

• The returned template shows <+input> placeholders that need values.

2. Choose input strategy

• Simple variables: pass flat key-value inputs (for example {"branch":"main","env":"prod"}).

• Complex/structural inputs: use input_set_ids (CI codebase/build blocks and nested template inputs are best handled this way).

• CI codebase shorthand keys (pipeline run only):

  Shorthand key	Expanded structure
  branch	build.type=branch, build.spec.branch=<value>
  tag	build.type=tag, build.spec.tag=<value>
  pr_number	build.type=PR, build.spec.number=<value>
  commit_sha	build.type=commitSha, build.spec.commitSha=<value>

• Constraint: shorthand expansion is skipped when inputs.build is already present (explicit build wins).

3. Execute the run

• harness_execute(resource_type="pipeline", action="run", resource_id="<pipeline_id>", ...)

• For Git-backed pipelines whose YAML should be loaded from a non-default branch, pass params.pipeline_branch (sent to Harness as pipelineBranchName):

  {
    "resource_type": "pipeline",
    "action": "run",
    "resource_id": "deploy_app",
    "params": { "pipeline_branch": "feature/new-stage" },
    "inputs": { "branch": "main" },
    "wait": true
  }

4. Optional: combine both

• Use input_set_ids for the base shape and inputs for simple overrides.

If required fields are unresolved, the tool returns a pre-flight error with expected keys and suggested input sets. You can inspect available shorthand mappings with harness_describe(resource_type="pipeline") (executeActions.run.inputShorthands).

Dynamic Pipeline Execution

Use pipeline_dynamic_execution.run when an agent or external system generates the full v0 pipeline YAML at runtime and needs to run it against an existing Harness pipeline shell. This is not a replacement for normal pipeline.run: the saved v0 pipeline must already exist, account-level and pipeline-level Allow Dynamic Execution must be enabled, and the caller needs Edit plus Execute permissions on the pipeline.

{
  "resource_type": "pipeline_dynamic_execution",
  "action": "run",
  "resource_id": "deploy_app",
  "body": {
    "yaml": "pipeline:\n  identifier: deploy_app\n  name: Deploy App\n  stages: []"
  },
  "params": {
    "module_type": "CD",
    "notes": "agent-generated dynamic run",
    "notify_only_user": true
  }
}

Constraints:

• body must be an object with a yaml field. Raw string bodies are rejected by the public harness_execute schema.

• body.yaml may be a YAML string or a JSON pipeline object; JSON is serialized to YAML before the request.

• Runtime <+input> placeholders are not resolved by this API. Submit fully resolved YAML.

• Input sets, selective stage execution, retry, and triggers are not supported by the dynamic execution endpoint.

• The action is high_write and uses the normal confirmation/auto-approval path. The response projects the API envelope to { "execution_id": "...", "status": "..." } and includes an openInHarness execution link when scope data is available.

If Harness rejects the run as not enabled, check both the account-level Allow Dynamic Execution setting and the pipeline-level toggle under Pipeline -> Advanced Options -> Dynamic Execution Settings.

Execution Input Forensics

Use execution_inputs after a run to inspect the merged input YAML that produced a specific execution. This is useful when a failure depends on input-set merging, Git-backed input set branches, or trigger/runtime values that are hard to reconstruct from the execution page alone.

{
  "resource_type": "execution_inputs",
  "resource_id": "PLAN_EXECUTION_ID",
  "params": {
    "resolve_expressions": true,
    "resolve_expressions_type": "RESOLVE_ALL_EXPRESSIONS"
  }
}

The get response is projected to:

• executionId - the plan execution ID from resource_id.

• inputSetYaml - merged runtime input YAML used for the run, or null.

• inputSetTemplateYaml - input template at execution time, or null.

• resolvedYaml - expression-resolved YAML when resolve_expressions=true, otherwise usually null.

• inputSetDetails - contributing saved input sets as { identifier, name } pairs.

• inputSetBranchName - source branch for Git-backed input sets, or null.

execution_inputs is get-only and read-risk. If resolve_expressions is omitted, the server omits the API query parameters and Harness uses its default UNKNOWN resolution mode.

Pipeline Execute Wait Mode

For pipeline.run, pipeline.retry, and pipeline_v1.run, pass wait: true to let the server poll until the execution reaches a terminal status. This keeps a pipeline launch and status check in one tool call instead of asking the client or LLM to run a polling loop.

{
  "resource_type": "pipeline",
  "action": "run",
  "resource_id": "deploy_app",
  "inputs": { "branch": "main" },
  "wait": true,
  "wait_timeout_seconds": 900,
  "wait_poll_interval_seconds": 5
}

Wait mode behavior:

• Default timeout is 600 seconds; allowed range is 10 seconds to 7200 seconds.

• Initial poll interval defaults to 3 seconds, backs off by 1.5x, and caps at 30 seconds.

• On success or failure, the response includes fields such as execution_id, execution_status, execution_terminal, execution_elapsed_ms, and execution_poll_count.

• If the timeout fires, the original trigger still succeeded; the response includes execution_timed_out: true and _wait.hint with the last observed status.

• If polling fails after the trigger succeeds, the response includes _wait.error and a recheck hint. Do not blindly rerun the pipeline unless you have confirmed the first execution is not running.

• Failed terminal statuses include _diagnose_hint pointing to harness_diagnose(resource_type="execution", options={execution_id: "..."}).

Ask the AI DevOps Agent to create a pipeline:

{
  "prompt": "Create a pipeline that builds a Go app with Docker and deploys to Kubernetes",
  "action": "CREATE_PIPELINE"
}

Update a service via natural language:

{
  "prompt": "Add a sidecar container for logging",
  "action": "UPDATE_SERVICE",
  "conversation_id": "prev-conversation-id",
  "context": [{ "type": "yaml", "payload": "<existing service YAML>" }]
}

Pipeline Storage Modes

Harness pipelines can be stored in three ways:

Create an inline pipeline (default):

// harness_create
{
  "resource_type": "pipeline",
  "body": {
    "yamlPipeline": "pipeline:\n  name: My Pipeline\n  identifier: my_pipeline\n  stages:\n    - stage:\n        name: Build\n        type: CI\n        spec:\n          execution:\n            steps:\n              - step:\n                  type: Run\n                  name: Echo\n                  spec:\n                    command: echo hello"
  }
}

Create a remote pipeline (External Git — e.g. GitHub):

// harness_create
{
  "resource_type": "pipeline",
  "body": {
    "yamlPipeline": "pipeline:\n  name: Deploy Service\n  identifier: deploy_service\n  stages: []"
  },
  "params": {
    "store_type": "REMOTE",
    "connector_ref": "my_github_connector",
    "repo_name": "my-repo",
    "branch": "main",
    "file_path": ".harness/deploy-service.yaml",
    "commit_msg": "Add deploy pipeline via MCP"
  }
}

Create a remote pipeline (Harness Code — no connector needed):

// harness_create
{
  "resource_type": "pipeline",
  "body": {
    "yamlPipeline": "pipeline:\n  name: Build App\n  identifier: build_app\n  stages: []"
  },
  "params": {
    "store_type": "REMOTE",
    "is_harness_code_repo": true,
    "repo_name": "product-management",
    "branch": "main",
    "file_path": ".harness/build-app.yaml",
    "commit_msg": "Add build pipeline via MCP"
  }
}

Update a remote pipeline:

// harness_update
{
  "resource_type": "pipeline",
  "resource_id": "deploy_service",
  "body": {
    "yamlPipeline": "pipeline:\n  name: Deploy Service\n  identifier: deploy_service\n  stages:\n    - stage:\n        name: Deploy\n        type: Deployment"
  },
  "params": {
    "store_type": "REMOTE",
    "connector_ref": "my_github_connector",
    "repo_name": "my-repo",
    "branch": "main",
    "file_path": ".harness/deploy-service.yaml",
    "commit_msg": "Update deploy pipeline via MCP",
    "last_object_id": "abc123",
    "last_commit_id": "def456"
  }
}

Import a pipeline from an external Git repo:

// harness_execute
{
  "resource_type": "pipeline",
  "action": "import",
  "params": {
    "connector_ref": "my_github_connector",
    "repo_name": "my-repo",
    "branch": "main",
    "file_path": ".harness/existing-pipeline.yaml"
  },
  "body": {
    "pipeline_name": "Existing Pipeline",
    "pipeline_description": "Imported from GitHub"
  }
}

Import a pipeline from a Harness Code repo:

// harness_execute
{
  "resource_type": "pipeline",
  "action": "import",
  "params": {
    "is_harness_code_repo": true,
    "repo_name": "product-management",
    "branch": "main",
    "file_path": ".harness/existing-pipeline.yaml"
  },
  "body": {
    "pipeline_name": "Existing Pipeline"
  }
}

Create a connector:

{
  "resource_type": "connector",
  "body": { "connector": { "name": "My Docker Hub", "identifier": "my_docker", "type": "DockerRegistry" } }
}

Delete a trigger:

{
  "resource_type": "trigger",
  "resource_id": "nightly-trigger",
  "pipeline_id": "my-pipeline"
}

List input sets for a pipeline:

{
  "resource_type": "input_set",
  "pipeline_id": "my-pipeline"
}

Get a specific input set:

{
  "resource_type": "input_set",
  "resource_id": "prod-inputs",
  "pipeline_id": "my-pipeline"
}

Create an input set:

{
  "resource_type": "input_set",
  "pipeline_id": "my-pipeline",
  "body": "inputSet:\n  name: Production Inputs\n  identifier: prod_inputs\n  pipeline:\n    identifier: my-pipeline\n    variables:\n      - name: env\n        type: String\n        value: production"
}

Update an input set:

{
  "resource_type": "input_set",
  "resource_id": "prod_inputs",
  "pipeline_id": "my-pipeline",
  "body": "inputSet:\n  name: Production Inputs\n  identifier: prod_inputs\n  pipeline:\n    identifier: my-pipeline\n    variables:\n      - name: env\n        type: String\n        value: production\n      - name: replicas\n        type: String\n        value: \"3\""
}

Delete an input set:

{
  "resource_type": "input_set",
  "resource_id": "prod_inputs",
  "pipeline_id": "my-pipeline"
}

Resource Types

218 resource types organized across 38 toolsets. Each resource type supports a subset of CRUD operations and optional execute actions.

Platform

Pipelines

Only one pipeline YAML resource type is loaded at startup. By default HARNESS_PIPELINE_VERSION=0 exposes pipeline and hides pipeline_v1; set HARNESS_PIPELINE_VERSION=1 to expose pipeline_v1 and hide pipeline. In HTTP mode, include x-harness-pipeline-version: 0 or 1 on the initialize request to choose the version for that session.

AI Agents

Services

Environments

Connectors

Infrastructure

Secrets

Execution Logs

Audit Trail

Delegates

Code Repositories

commit creation commits one or more file actions directly through the Harness Code API without cloning. Pass body.title, body.branch, and body.actions; each action is CREATE, UPDATE, DELETE, or MOVE, and UPDATE requires the current blob SHA.

Artifact Registries

File Store

file_store manages Harness File Store files and folders through the generic tools. It supports account, org, and project scope; pass resource_scope="account"|"org"|"project" or paste a Harness File Store URL so the server can derive scope and IDs.

Common calls:

# List the account-level File Store.
harness_list(resource_type="file_store", resource_scope="account")

# Create a folder at the current scope root.
harness_create(resource_type="file_store", body={
  name: "scripts",
  type: "FOLDER",
  parent_identifier: "Root"
})

# Upload a UTF-8 script file. Use content_base64 instead for binary data.
harness_create(resource_type="file_store", body={
  name: "deploy.sh",
  type: "FILE",
  parent_identifier: "Root",
  content: "#!/usr/bin/env bash\n./deploy",
  mime_type: "text/x-shellscript",
  file_usage: "SCRIPT"
})

# Rename metadata without replacing file content.
harness_update(resource_type="file_store", resource_id="deploy_script", body={
  name: "deploy-prod.sh",
  type: "FILE",
  parent_identifier: "Root"
})

# List first-level children of a folder. This is a read-risk execute action.
harness_execute(resource_type="file_store", action="list_children",
  resource_id="scripts_folder", params={folder_name: "scripts"})

Multipart body constraints:

• Create/update accept JSON body, then convert it to multipart/form-data for /ng/api/file-store.

• name, type (FILE or FOLDER), and parent_identifier are required; use the literal "Root" only for the root of the selected scope.

• FILE create requires exactly one of content (UTF-8 string) or content_base64 (valid non-empty base64). FILE update can omit content for metadata-only updates, or provide exactly one content field to replace content.

• FOLDER create/update must omit content and content_base64.

• Optional file_usage must be MANIFEST_FILE, CONFIG, or SCRIPT; optional scalar metadata such as description, mime_type, path, and tags must be strings.

• Upload content is capped at 100 MB. Confirmation prompts redact content, content_base64, and contentBase64 previews before elicitation.

list_children accepts either shorthand (resource_id plus params.folder_name, or params.file_store_id/params.folder_identifier plus params.folder_name) or a full FileStoreNode body with identifier, name, and type: "FOLDER". Full bodies use Harness camelCase parentIdentifier; shorthand may use params.parent_identifier.

Templates

Template operations use the Harness Template service paths (/template/api/templates...). Create and update require the full template YAML string in body.template_yaml or body.yaml; version_label targets a specific version for update/delete, while deleting without version_label deletes all versions.

Dashboards

Database DevOps

Infrastructure as Code Management (IaCM)

IaCM resources are default-enabled and mostly project-scoped. Start with iacm_workspace to find workspace identifiers, then use that workspace_id for workspace resources, costs, and activity diffs. The module registry is account-scoped.

Typical workflow:

1. harness_list(resource_type="iacm_workspace", org_id="...", project_id="...") to find the workspace.

2. harness_list(resource_type="iacm_resource", org_id="...", project_id="...", workspace_id="...") to inspect Terraform resources, outputs, and data sources.

3. harness_list(resource_type="iacm_workspace_costs", org_id="...", project_id="...", workspace_id="...") to review per-execution cost entries.

4. harness_list(resource_type="iacm_activity_resource_change", org_id="...", project_id="...", activity_id="...", workspace_id="...") to inspect before/after resource diffs for a plan, apply, or destroy activity.

IaCM list responses expose page_count as the count for the current page only. When has_more is true, keep requesting the next 1-based page and sum page counts if you need a total.

Internal Developer Portal (IDP)

Pull Requests

Use harness_execute(resource_type="pull_request", action="close", ...) for an explicit close operation. harness_update also accepts body.state (open or closed) and routes state changes to the dedicated Harness Code PR state endpoint; send title/description edits in a separate update call.

Feature Flags

FME (Split.io) resources — fme_* resources use the Split.io API (api.split.io) and are scoped by workspace ID rather than org/project. In single-user/self-hosted mode, auth uses a Bearer token from HARNESS_FME_API_KEY, falling back to a non-placeholder HARNESS_API_KEY. HARNESS_FME_API_KEY may be a legacy Split admin key or an FME-entitled Harness PAT/SAT, but it is rejected in multi-user mode so shared deployments cannot override each session user's credential. Hosted OAuth/service-routing credentials for Harness platform APIs do not authenticate direct Split.io requests. fme_feature_flag supports full lifecycle management: create (requires traffic_type_id), list, get, update metadata, delete, and kill/restore/archive/unarchive execute actions. Use fme_traffic_type to discover traffic type IDs, fme_identity to create/update identity attributes, and fme_standard_segment / fme_segment_keys to inspect standard segments and add member keys. fme_rule_based_segment provides CRUD for targeting segments, while fme_rule_based_segment_definition manages environment-specific segment rules with enable/disable and change request approval flows.

GitOps

Chaos Engineering

Cloud Cost Management (CCM)

Software Engineering Insights (SEI)

SEI resources are consolidated for token efficiency. Use metric or aspect params for DORA, team/org-tree details, and AI insights.

Software Supply Chain Assurance (SCS)

Security Testing Orchestration (STO)

security_exemption create is a high_write operation. The server derives requester_id from the authenticated PAT, sets exemptFutureOccurrences=true, and defaults duration_days to 30 when not provided. For listing exemptions, pass a small explicit page size (for example filters: { "status": "Pending", "size": 5 }) and follow the _nextPageHint returned in each response.

Security exemption execute workflow:

• Use harness_list with resource_type="security_exemption" and an explicit status such as Pending, Approved, Rejected, Expired, or Canceled.

• Use harness_execute with action="approve" and a required body.scope: CURRENT, ACCOUNT, ORG, or PROJECT. CURRENT approves at the exemption's existing scope; the other scopes use the STO promote endpoint internally. The server auto-fills body.approver_id from the authenticated user when omitted; body.comment is optional.

• Use action="reject" to reject an exemption. body.approver_id is also auto-filled when omitted.

• There is no separate promote execute action. Use action="approve" with a non-CURRENT body.scope when the requested outcome is approval at account, organization, or project scope.

Access Control

Governance

Deployment Freeze

Service Overrides

Settings

Visualizations

Inline PNG chart visualizations rendered from Harness data. These are metadata-only resource types with no API operations — they exist so the LLM can discover available chart types via harness_describe. Use include_visual=true on supported tools (harness_diagnose, harness_list, harness_status) to generate charts.

MCP Prompts

DevOps

FinOps

DevSecOps

Harness Code

MCP Resources

Toolset Filtering

By default, 38 of 39 toolsets are enabled. One toolset is opt-in and excluded from the defaults:

• ansible — Harness Ansible (inventories, playbooks, hosts, activity). Opt-in because it is project-scoped and adds concepts many users do not need.

Adding toolsets with + prefix

Use the + prefix to explicitly include opt-in toolsets alongside all defaults:

# Explicitly include Ansible alongside all defaults
HARNESS_TOOLSETS=+ansible

Removing default toolsets

Use the - prefix to exclude toolsets you don't need:

# Remove chaos and ccm from defaults
HARNESS_TOOLSETS=-chaos,-ccm

Combining + and -

# Add Ansible, remove chaos
HARNESS_TOOLSETS=+ansible,-chaos

Explicit allowlist

An explicit comma-separated list (no prefixes) replaces the defaults entirely. Only the listed toolsets are enabled:

# Only expose pipelines, services, and connectors
HARNESS_TOOLSETS=pipelines,services,connectors

Available toolset names:

Architecture

                 +------------------+
                 |   AI Agent       |
                 |  (Claude, etc.)  |
                 +--------+---------+
                          |  MCP (stdio or HTTP)
                 +--------v---------+
                |    MCP Server     |
                | 11 Generic Tools  |
                 +--------+---------+
                          |
                 +--------v---------+
                |    Registry       |  <-- Declarative resource definitions
                |  38 Toolsets      |      (data files, not code)
                |  218 Resource Types|
                 +--------+---------+
                          |
                 +--------v---------+
                 |  HarnessClient    |  <-- Auth, retry, rate limiting
                 +--------+---------+
                          |  HTTPS
                 +--------v---------+
                 |  Harness REST API |
                 +-------------------+

How It Works

1. Tools are generic verbs: harness_list, harness_get, etc. They accept a resource_type parameter that routes to the correct API endpoint.

2. The Registry maps each resource_type to a ResourceDefinition — a declarative data structure specifying the HTTP method, URL path, path/query parameter mappings, and response extraction logic.

3. Dispatch resolves the resource definition, builds the HTTP request (path substitution, query params, resource_scope-aware account/org/project injection), calls the Harness API through HarnessClient, and extracts the relevant response data.

4. Toolset filtering (HARNESS_TOOLSETS) controls which resource definitions are loaded into the registry at startup.

5. Structured output is declared with MCP outputSchema; harness_list coerces arrays and common list wrappers into object-shaped structuredContent for strict clients.

6. Deep links are automatically appended to responses, providing direct Harness UI URLs for every resource.

7. Compact mode strips verbose metadata from list results, keeping only actionable fields (identity, status, type, timestamps, deep links) to minimize token usage.

Adding a New Resource Type

Create a new file in src/registry/toolsets/ or add a resource to an existing toolset:

// src/registry/toolsets/my-module.ts
import type { ToolsetDefinition } from "../types.js";

export const myModuleToolset: ToolsetDefinition = {
  name: "my-module",
  displayName: "My Module",
  description: "Description of the module",
  resources: [
    {
      resourceType: "my_resource",
      displayName: "My Resource",
      description: "What this resource represents",
      toolset: "my-module",
      scope: "project",                    // "project" | "org" | "account"
      identifierFields: ["resource_id"],
      listFilterFields: ["search_term"],
      operations: {
        list: {
          method: "GET",
          path: "/my-module/api/resources",
          queryParams: { search_term: "search", page: "page", size: "size" },
          responseExtractor: (raw) => raw,
          description: "List resources",
        },
        get: {
          method: "GET",
          path: "/my-module/api/resources/{resourceId}",
          pathParams: { resource_id: "resourceId" },
          responseExtractor: (raw) => raw,
          description: "Get resource details",
        },
      },
    },
  ],
};

Then import it in src/registry/index.ts and add it to the ALL_TOOLSETS array. No changes needed to any tool files.

Development

# Build
pnpm build

# Watch mode
pnpm dev

# Type check
pnpm typecheck

# Run tests
pnpm test

# Watch tests
pnpm test:watch

# Interactive MCP Inspector
pnpm inspect

# Refresh generated README counts from the built registry
pnpm docs:generate

# Verify README counts and clone instructions are current
pnpm docs:check

# Sync and verify JSON Schemas used by harness_schema
pnpm sync-schemas
pnpm check-schema-coverage

Project Structure

src/
  index.ts                          # Entrypoint, transport setup
  config.ts                         # Env var validation (Zod)
  client/
    harness-client.ts               # HTTP client (auth, retry, rate limiting)
    types.ts                        # Shared API types
  registry/
    index.ts                        # Registry class + dispatch logic
    types.ts                        # ResourceDefinition, ToolsetDefinition, etc.
    toolsets/                        # One file per toolset (declarative data)
      platform.ts
      pipelines.ts
      services.ts
      ccm.ts
      access-control.ts
      ...
  tools/                            # 11 generic MCP tools
    harness-list.ts
    harness-get.ts
    harness-create.ts
    harness-update.ts
    harness-delete.ts
    harness-execute.ts
    harness-search.ts
    harness-diagnose.ts
    harness-describe.ts
    harness-status.ts
    harness-schema.ts

  resources/                        # MCP resource providers
    pipeline-yaml.ts
    execution-summary.ts
  prompts/                          # MCP prompt templates
    build-deploy-app.ts             # DevOps: end-to-end build & deploy workflow
    debug-pipeline.ts               # DevOps: debug failed executions
    create-pipeline.ts              # DevOps: generate pipeline from requirements
    onboard-service.ts              # DevOps: onboard new service
    dora-metrics.ts                 # DevOps: DORA metrics review
    setup-gitops.ts                 # DevOps: GitOps application setup
    chaos-resilience.ts             # DevOps: chaos experiment design
    feature-flag-rollout.ts         # DevOps: progressive flag rollout
    migrate-to-template.ts          # DevOps: extract templates from pipeline
    delegate-health.ts              # DevOps: delegate health check
    developer-scorecard.ts          # DevOps: IDP scorecard review
    optimize-costs.ts               # FinOps: cost optimization
    cloud-cost-breakdown.ts         # FinOps: cost deep-dive
    commitment-utilization.ts       # FinOps: RI/savings plan analysis
    cost-anomaly.ts                 # FinOps: anomaly investigation
    rightsizing.ts                  # FinOps: rightsizing recommendations
    security-review.ts              # DevSecOps: security issue review
    vulnerability-triage.ts         # DevSecOps: vulnerability triage
    sbom-compliance.ts              # DevSecOps: SBOM compliance audit
    supply-chain-audit.ts           # DevSecOps: supply chain audit
    exemption-review.ts             # DevSecOps: exemption approval
    access-control-audit.ts         # DevSecOps: access control audit
    code-review.ts                  # Harness Code: PR code review
    pr-summary.ts                   # Harness Code: auto-generate PR summary
    branch-cleanup.ts               # Harness Code: stale branch cleanup
    pending-approvals.ts            # Approvals: find and act on pending approvals
  utils/
    cli.ts                          # CLI arg parsing (transport, port)
    errors.ts                       # Error normalization
    logger.ts                       # stderr-only logger
    progress.ts                     # MCP progress & logging notifications
    rate-limiter.ts                 # Client-side rate limiting
    deep-links.ts                   # Harness UI deep link builder
    response-formatter.ts           # Consistent MCP response formatting
    compact.ts                      # Compact list output for token efficiency
tests/
  config.test.ts                    # Config schema validation tests
  utils/
    response-formatter.test.ts
    deep-links.test.ts
    errors.test.ts
  registry/
    registry.test.ts                # Registry loading, filtering, dispatch tests

Elicitation

The write tools (harness_create, harness_update, harness_delete, harness_execute) use MCP elicitation to prompt the user for confirmation when the action's risk requires it — medium_write, high_write, and destructive operations only. Low-risk creates / updates / reads (e.g. pipeline.create, pipeline.update, hql_query.run) proceed silently with no prompt. When a prompt is surfaced, the user sees what's about to happen and accepts or declines, giving real human-in-the-loop approval for the operations that actually mutate or run things.

How it works:

1. The LLM calls a write tool with medium_write+ risk (e.g. harness_delete, harness_execute pipeline.run). Low-risk creates / updates / reads do not surface a prompt.

2. The server sends an elicitation request to the client with a summary of the operation and a confirm checkbox (default checked).

3. The user sees the details and clicks Accept (with confirm checked) or Decline / Cancel.

4. If accepted with confirm: true, the operation proceeds. If accepted with confirm unchecked, declined, or cancelled, it's blocked and the LLM is told (an explicit decline is authoritative and not bypassed by confirm: true on the tool call).

Client support:

Elicitation behavior varies by operation risk when client support is missing:

If elicitInput fails at runtime (transport error, unsupported method) for a medium_write+ operation, the call is blocked unless the caller passes confirm: true. confirm: true is honored as a fallback when the client could not surface a prompt or returned a degenerate accept ({action: "accept"} without the confirm field), but it does not override an explicit decline/cancel from a client that completed the elicitation handshake.

Autonomous Mode

Autonomous mode means the server proceeds with all operations — including writes and destructive actions — without prompting for confirmation. Enable it by setting:

HARNESS_AUTO_APPROVE_RISK=all

This is the deployment-level ceiling: once set, individual sessions cannot escalate beyond it (though they can choose a stricter threshold per-session via the x-harness-auto-approve-risk header).

Or in your MCP client config:

{
  "mcpServers": {
    "harness": {
      "command": "npx",
      "args": ["harness-mcp-v2"],
      "env": {
        "HARNESS_API_KEY": "pat.xxx.xxx.xxx",
        "HARNESS_AUTO_APPROVE_RISK": "all"
      }
    }
  }
}

Partial autonomy: You can also auto-approve only up to a specific risk level while still prompting for higher-risk operations:

# Auto-approve reads and low-risk writes; prompt for medium_write, high_write, destructive
HARNESS_AUTO_APPROVE_RISK=low_write

# Auto-approve up to high-risk writes; only prompt for destructive operations
HARNESS_AUTO_APPROVE_RISK=high_write

Autonomous mode warning: HARNESS_AUTO_APPROVE_RISK=all skips confirmation for all operations including harness_delete. Use with caution and consider pairing with HARNESS_TOOLSETS to restrict which resource types are available.

Migration note: HARNESS_SKIP_ELICITATION=true is still supported and maps to HARNESS_AUTO_APPROVE_RISK=all. A deprecation warning is logged to stderr. If both are set, HARNESS_AUTO_APPROVE_RISK takes precedence.

Safety

• Secrets are never exposed. The secret resource type returns metadata only (name, type, scope) — secret values are never included in any response.

• Confirmation-requiring operations use elicitation when available. When a write or execute action has medium_write, high_write, or destructive risk, harness_create, harness_update, harness_delete, and harness_execute attempt MCP elicitation before proceeding (see Elicitation). Low-risk actions (read, low_write — e.g. pipeline.create, pipeline.update, hql_query.run) proceed silently with no prompt.

• Medium-risk and above fail closed. If confirmation cannot be obtained for medium_write, high_write, or destructive operations, they are blocked instead of executing blindly. Override with HARNESS_AUTO_APPROVE_RISK for autonomous workflows.

• CORS restricted to same-origin. The HTTP transport only allows same-origin requests, preventing CSRF attacks from malicious websites targeting the MCP server on localhost.

• HTTP rate limiting. The HTTP transport enforces 60 requests per minute per IP to prevent request flooding.

• API rate limiting. The Harness API client enforces a 10 requests/second limit to avoid hitting upstream rate limits.

• Pagination bounds enforced. List queries are capped at 10,000 items total and 100 per page to prevent memory exhaustion.

• Retries with backoff. Transient failures (HTTP 429, 5xx) are retried with exponential backoff and jitter.

• Localhost binding. The HTTP transport binds to 127.0.0.1 by default — not accessible from the network.

• No stdout logging. All logs go to stderr to avoid corrupting the stdio JSON-RPC transport.

Complementary Skills

The Harness MCP server pairs well with Harness Skills — a collection of ready-made Claude Code skills (slash commands) designed for common Harness workflows. Install them alongside this MCP server to get high-level automation like /deploy, /rollback, /triage, and more without writing custom prompts.

Troubleshooting & Common Pitfalls

License

MIT