Harness MCP server: 11 tools, 240 resources

How Harness covers a whole delivery platform with eleven tools, what that design costs an agent's context, and where the cost moves to rather than disappearing.

Last verified 1 September 2026 · from Harness's repository, its documentation and the server's own tools/list response

This page is one of 90 in a directory of monitoring and developer tools. Each page checks what an AI coding agent can get out of the tool through its MCP server, the connector that lets the agent query the tool directly.

What is the Harness MCP server?

It is Harness's server for the Model Context Protocol. Harness is a delivery platform, so the agent gains reach across pipelines, services, environments, feature flags and cost data.

The eleven tools are verbs rather than nouns: harness_list, harness_get, harness_create, harness_update, harness_delete, harness_execute, harness_search, harness_describe, harness_status, harness_diagnose and harness_schema. The resource type is an argument.

Harness states the reason plainly: the model picks from eleven tools instead of hundreds. Several tools also accept a Harness URL and read the account, organization and project scope out of it.

Harness covers release broadly, reaches part of the way into code review through security testing and into cost through cloud cost management, and reaches nothing in agent observability.

For an agent, Harness answers questions about how software gets delivered, from the pipeline to the flag to the cloud bill. An agent connected here can find the failing pipeline stage. It can check which feature flag is on. It can read what the deployment cost. It cannot tell you what the running service is doing, so this connector pairs with a telemetry one.

What can an AI agent do with Harness?

Consolidation changes what an allowlist can do. On most connectors an allowlist can approve one tool and refuse another. Here harness_execute is one entry. It covers pipeline runs, flag kills, GitOps syncs and chaos experiments. The permission boundary has to be the API key's scope instead.

What connecting costs before the first question

A context window is the amount of text a model can hold at once, and connecting uses some of it before the agent asks anything. Every tool the server registers sends its name, description and input schema to the model when the session starts. The tool list is charged whether or not a single tool is called.

Eleven tools cost about 8,625 tokens. Playwright registers twenty-four tools for less, and Arize Phoenix twenty-seven. Consolidation moved the cost rather than removing it.

The schemas carry it: 3,230 characters per tool, because a tool that can reach 240 resource types has to describe what it accepts for them. Descriptions are short, at 282 characters, so this is structure rather than prose.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted 24 August 2026 from tools/list on the released v3.2.20 bundle's Node server. It was run locally with a placeholder key.

Registration is static, so no Harness account was involved. Figures cover name, description and input schema. The two models in column one share the o200k_base encoding, so they get one count. The spread is formatting, about 43%. Reasoning tokens are output, spent per request. This count is input, charged once at connection.

The Claude figures are measured rather than encoded. Anthropic publishes no local tokenizer. The definitions were sent to the API's count_tokens endpoint, and a no-tools baseline was subtracted. The range covers two tokenizer generations rather than two serializations. Haiku 4.5 and Opus 4.6 sit at the low end. Opus 5 and Sonnet 5 sit at the high end. The gap between them is about a third.

Harness MCP server limits: where answers come back incomplete

The limits below follow from the dispatch design rather than from anything failing.

How to configure Harness MCP for agents

1. 1

   Scope the API key, not the tool list

   With harness_create, harness_update, harness_delete and harness_execute covering 240 resource types, an allowlist cannot separate a pipeline retry from a feature-flag kill. The key's permissions are the real boundary.

2. 2

   Set the default scope

   HARNESS_DEFAULT_ORG_ID and HARNESS_DEFAULT_PROJECT_ID stop the agent querying the wrong project. PAT means personal access token, and SAT means service account token. Both carry the account ID, so HARNESS_ACCOUNT_ID is only needed for keys without one.

3. 3

   Filter toolsets if you do not need everything

   39 toolsets are on by default. Harness does not publish their names in the README, so this page cannot list them; Ansible coverage is the one named as opt-in.

4. 4

   Protect any shared deployment

   HARNESS_MCP_AUTH_TOKEN makes every request to /mcp require a bearer token, which matters the moment the server is reachable by anything but you.

Paste this into your agent’s instructions

Dispatch hides which resource an action touched, and scope errors return empty rather than failing. The block makes the agent name both.

When you answer using Harness MCP tools, state:
- The account, organization and project scope you queried.
- The resource type you passed, not just the tool name. A
  harness_get on the wrong type returns nothing.
- Whether you called harness_execute, and on what. It runs
  pipelines, kills feature flags and starts chaos experiments.
Never state that data does not exist. State that none was
returned under the scope above.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

Sometimes. Harness has a documented REST API, and for one known operation a direct call is cheaper than 8,625 tokens of tool definitions.

The connector earns its place on breadth. An agent that does not know which of 240 resource types holds the answer can use harness_search and harness_describe to find out, which is not a thing the REST API makes easy.

The hosted MCP service is the third option. It authenticates with Harness Platform OAuth rather than an API key, and Harness Support has to enable it per account.

Connect the server when the agent is exploring the platform. Call the API directly when you already know the endpoint.

Harness MCP server setup

The package runs with npx and needs a Harness API key. PAT and SAT tokens carry the account ID, so most setups need nothing else.

There is also a hosted service, which authenticates with Harness Platform OAuth rather than an API key and has to be enabled per account by Harness Support.

claude mcp add-json "harness" '{"command":"npx",
  "args":["-y","harness-mcp-v2","stdio"],
  "env":{"HARNESS_API_KEY":"pat.<account>.<id>.<secret>"}}'

{"mcpServers":{"harness":{"command":"npx",
  "args":["-y","harness-mcp-v2","stdio"],
  "env":{"HARNESS_API_KEY":"<your-key>",
         "HARNESS_DEFAULT_ORG_ID":"<org>",
         "HARNESS_DEFAULT_PROJECT_ID":"<project>"}}}}

[mcp_servers.harness]
command = "npx"
args = ["-y", "harness-mcp-v2", "stdio"]
env = { HARNESS_API_KEY = "<your-key>" }

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

{"mcpServers":{"harness":{"command":"npx",
  "args":["-y","harness-mcp-v2","stdio"],
  "env":{"HARNESS_API_KEY":"<your-key>","HARNESS_DEFAULT_ORG_ID":"<org>","HARNESS_DEFAULT_PROJECT_ID":"<project>"}}}}

{"mcpServers":{"harness":{"command":"npx",
  "args":["-y","harness-mcp-v2","stdio"],
  "env":{"HARNESS_API_KEY":"<your-key>","HARNESS_DEFAULT_ORG_ID":"<org>","HARNESS_DEFAULT_PROJECT_ID":"<project>"}}}}

{"servers":{"harness":{"command":"npx",
  "args":["-y","harness-mcp-v2","stdio"],
  "env":{"HARNESS_API_KEY":"<your-key>","HARNESS_DEFAULT_ORG_ID":"<org>","HARNESS_DEFAULT_PROJECT_ID":"<project>"}}}}

{"mcpServers":{"harness":{"command":"npx",
  "args":["-y","harness-mcp-v2","stdio"],
  "env":{"HARNESS_API_KEY":"<your-key>","HARNESS_DEFAULT_ORG_ID":"<org>","HARNESS_DEFAULT_PROJECT_ID":"<project>"}}}}

HARNESS_API_KEY=<your-key> \
HARNESS_MCP_AUTH_TOKEN=<shared-secret> \
npx -y harness-mcp-v2 http --port 8080

Read-only access, permissions and security

• The API key is the permission boundary

  harness_create, harness_update, harness_delete and harness_execute reach 240 resource types, so a client allowlist cannot express "read pipelines but do not delete connectors". Scope the key instead.

• One tool executes across the platform

  harness_execute runs and interrupts pipelines, kills and restores feature flags, syncs GitOps apps and runs chaos experiments.

• A shared deployment is open by default

  HARNESS_MCP_AUTH_TOKEN adds bearer authentication to /mcp; without it, any client that can reach the port acts with the server's key.

• The hosted service authenticates differently

  It uses Harness Platform OAuth, and Harness Support must enable it per account before the endpoint works.

Troubleshooting

The server will not start

An API key without an embedded account segment needs HARNESS_ACCOUNT_ID. PAT and SAT tokens carry it; other key types do not.

A resource that exists is not found

Check the scope. Set HARNESS_DEFAULT_ORG_ID and HARNESS_DEFAULT_PROJECT_ID, or paste a Harness URL and let the tool read the scope from it.

Pipeline YAML is rejected

Call harness_schema for the resource type first. Pipeline and template schemas ship bundled with the server; others are fetched.

An action ran that you did not expect

harness_execute covers pipeline runs, flag kills, GitOps syncs and chaos experiments. Narrow the API key rather than the allowlist.

Harness MCP server: Reference

What engineers report

Harness's connector has a modest public footprint, so the useful signals are its design and its coverage.

Repository facts from the GitHub API, retrieved 24 August 2026. Tool counts come from the server's own tools/list response on the released bundle. The resource-type figure is Harness's own, from the repository README.

Should you connect an agent to Harness?

Yes. Best for harness teams scoping the API key tightly, since the key is the real permission boundary here.

Strongest fit

Harness teams who want one connection covering delivery, flags and cloud cost rather than three.

Main advantage

The broadest platform reach in this guide, from eleven tools an agent can hold in mind at once.

Main weakness

Consolidation moves cost into schemas, so eleven tools cost more than another vendor's twenty-seven.

Operational risk

A single execute tool spans pipeline runs, flag kills and chaos experiments, which an allowlist cannot separate.

Frequently asked questions