Base44
Build and edit Base44 apps from an AI assistant, and query data your app's users submit.
Run pipelines, toggle flags, trace failed deploys and read cloud costs.
Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.
Paste it into any MCP client. Setup docs
The Harness MCP server connects the Harness software delivery platform to Claude, ChatGPT and any MCP-compatible agent. Its 11 tools are generic verbs, not fixed features: Harness routes them to 228 resource types across CI/CD, GitOps, feature flags, cloud cost and chaos engineering. One small tool list runs pipelines, traces failed deployments and toggles flags.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
No. Harness states in its server repository that secret values are never included in any response, and the secret resource type supports only list and get, returning metadata such as name, type and scope. There is no create, update or delete for secrets through this connector, so an assistant can inventory your secrets but never read or change one.
Yes, if your own Harness permissions allow it. Running a pipeline is a harness_execute action, and a Harness pipeline can target any environment it is configured for, including production. Harness rates a pipeline run as high_write, so the connector asks you to confirm before it triggers, and your account's approval stages and RBAC still apply to the execution itself.
Because Harness does not express its permissions as OAuth scopes. Its authorization server at id.harness.io advertises exactly three scopes, all of them identity claims rather than capabilities. The real boundary is Harness RBAC, so the connector reaches whatever your roles and resource groups already permit, and the consent screen has no read-only option to pick.
Eleven tools reach 228 resource types, by Harness's own count. The write verbs take a resource type as an argument rather than being scoped to one entity, so harness_create, harness_update, harness_delete and harness_execute apply across pipelines, connectors, environments, services, feature flags, chaos experiments and more. Read Harness's resource tables to see which verbs each type supports.
Not reliably, and this is the important caveat. Harness uses MCP elicitation for medium-risk operations and above, but its own client table lists Claude Desktop as not yet supporting elicitation. Where a client cannot show the prompt, Harness blocks the operation rather than running it, unless the caller passes confirm true explicitly.
Not on the hosted endpoint. Harness ships a HARNESS_READ_ONLY environment variable that blocks create, update, delete and execute, but environment variables only apply to the server you run yourself. On the managed endpoint at mcp.harness.io, restriction comes from Harness RBAC on the identity you sign in with, so grant a narrow role instead.
Yes. Harness states the hosted MCP service must be enabled and configured per account by Harness Support before the endpoint can be used. The hosted service also authenticates with Harness Platform OAuth rather than an API key, and client configuration cannot change which Harness environment the managed endpoint routes to.
README.md — https://github.com/harness/mcp-server (fetched 2026-08-19, raw file). Verified as Harness's own repository: owner harness, described as the official repo, MIT licensed, last pushed 2026-08-19 · retrieved 2026-08-19src/prompts, 35 files, listed 2026-08-19) · retrieved 2026-08-19robots.txt — https://developer.harness.io/robots.txt (fetched 2026-08-19). Allow: / with only /docs/first-gen/ disallowed; no Content-Signal line · retrieved 2026-08-19llms.txt — https://developer.harness.io/llms.txt (fetched 2026-08-19) · retrieved 2026-08-19https://mcp.harness.io/.well-known/oauth-protected-resource and /.well-known/oauth-protected-resource/mcp, both HTTP 200 (2026-08-19). The two bodies differ only in resource, declaring the host and the /mcp path respectively; both name https://id.harness.io as the authorization server · retrieved 2026-08-19https://id.harness.io/.well-known/oauth-authorization-server, HTTP 200 (2026-08-19). scopes_supported present with organization, profile, email; code_challenge_methods_supported S256; a registration_endpoint is published · retrieved 2026-08-19https://mcp.harness.io/mcp returned HTTP 401 with no www-authenticate header (2026-08-19) · retrieved 2026-08-19Connect Harness once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.