Connect an AI agent with MCP

Add a remote MCP server in your agent's MCP settings with this URL. It speaks Streamable HTTP:

https://app.cachely.dev/mcp

The first call returns 401 with a WWW-Authenticate header that points the client at the OAuth metadata. A client that supports MCP authorization takes it from there: it registers itself, opens a browser window, and you sign in to Cachely with Google or GitHub.

It is standard OAuth 2.1 with dynamic client registration and PKCE (S256); there is no Cachely-specific grant and no API key to paste. The agent acts as you: it sees only the workspaces you can access, and the access token is refreshable, so you sign in once.

Discovery starts at /.well-known/oauth-protected-resource. The step-by-step flow, for clients that implement it themselves, is in auth.md, and the server is described for MCP registries by its /.well-known/mcp/server-card.json server card.

If you belong to more than one account, send an x-cachely-account header to choose which one the agent works in.

list_workspaces

List your remote-cache workspaces.

create_workspace

Create a workspace, named after the repository by default. Rejected with 403 when your plan's workspace limit is reached.

create_token

Mint a token for a workspace with a name, a scope (read, write, or read,write), and an optional expiry. The raw token is returned once. Pass usage as developer for a laptop or pull-request build and the response warns when that token can write. Rejected with 403 at the plan's token limit and 409 when the name is taken.

setup_instructions

Return the config files for Nx, Lerna, Turborepo, Gradle, or Bazel, as a dotenv file, shell exports, or a GitHub Actions secret, with the paths to gitignore. Read-only: the token in the output is always a placeholder, so the files are safe to commit.

insights_summary

Hit rate, hits, misses, uploads, bytes, and estimated build time saved over a look-back window (30 days by default, up to 365), filterable by workspace, project, and tool. Clients that support MCP Apps render it as a chart.

insights_projects and insights_tools

The same numbers broken down per project or per build tool.

get_plan_usage

Your plan, its limits, this period's usage, and what the next plan up would change.

The agent cannot read or download cached artifacts, delete a workspace, or revoke a token. Those stay in the dashboard.

Once connected, one request covers the whole setup:

Set up the Cachely remote cache for this repository: create a workspace
named after the repo, mint a read-write token for CI, write the config
from setup_instructions, and put the token only in a gitignored file.

A token is a secret. Ask for a read-only token for developer machines, and keep the raw value out of chat history and commits - see tokens and access control.

The MCP server does the account work. For reviewing the repository itself - which targets are cacheable, what makes a cache unsafe to share - install the Cachely agent skills or run the cache audit.