CircleCI MCP Server: Hosted vs CLI, Pricing, Risks, and Evaluation

Full Guide

Kingy verdict: CircleCI’s current MCP story is more useful—and more complicated—than the original launch record suggests. The standalone Node server from April 2025 is deprecated. New evaluations should compare CircleCI’s hosted MCP server for focused run diagnosis with the broader MCP interface built into the CircleCI CLI, then put explicit confirmation around every mutating or administrative action.

What the current product is

CircleCI now documents two independent MCP routes. The hosted server runs at CircleCI and connects over HTTPS using OAuth 2.0 or a personal API token. Its curated tool set follows the hierarchy of a CI run: list runs, inspect workflows and jobs, read failed-step logs, review test results and artifacts, then rerun or cancel a workflow. The CLI MCP runs locally and inherits credentials from the CircleCI CLI.

The CLI route is much broader. CircleCI says it exposes the CLI surface for configuration authoring and validation, projects and contexts, environment variables, policies, runner resources, code signing, deployments, cache operations and direct API calls. That extra reach makes it useful for platform engineering, but it also changes the risk profile. A diagnostic assistant and an administrative assistant should not receive the same credential or confirmation policy simply because both speak MCP.

The deprecation matters

CircleCI’s April 9, 2025 changelog announced a standalone open-source server. The current changelog and product documentation mark that implementation deprecated or in maintenance mode and direct new users to the hosted server or the CLI-integrated MCP. Existing CircleCI Server installations may still need the standalone package, but its repository should not be presented as the primary forward path.

This transition is also a migration check. Teams should inventory old npx configurations and long-lived personal tokens before adding a new server. Otherwise, an assistant can retain two overlapping CircleCI connections with different versions and credentials, making tool selection and audit attribution harder to reason about.

Pricing and operational cost

CircleCI does not publish a separate MCP add-on price on the reviewed pages. Access depends on the underlying CircleCI organization, plan, compute credits, users, storage and any applicable add-ons. The standalone repository is public, but open source does not make the CircleCI service or the agent’s model usage free. Estimate cost from the workloads the assistant can trigger, especially reruns, evaluation pipelines and administrative commands.

How Kingy would evaluate it

Start with hosted MCP and a read-oriented task pack: locate the latest failed run on a branch, identify the failed step, retrieve only the relevant log range, list failing tests and name available artifacts. Reconcile every answer with the CircleCI interface or API. Then test access boundaries across two projects, a fork, a private artifact and a user whose permissions were just revoked.

Evaluate mutations separately. Require the assistant to state the target workflow, branch, current status, intended action and expected cost before a rerun or cancellation. For the CLI MCP, use a dedicated low-privilege identity and block context, environment-variable, runner-token, signing and policy changes until each has an approved use case. Record the CircleCI actor and MCP client for every action.

Kingy reviewed CircleCI’s current product page, MCP overview, launch changelog, public repository and pricing material. We did not connect a CircleCI organization, inspect live logs or execute a workflow.

Primary sources

• CircleCI MCP overview and current server comparison
• CircleCI MCP product page
• Original launch changelog and deprecation notice
• Standalone MCP repository
• CircleCI pricing