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.

Creating a token through MCP returns its raw value to your coding agent. Keeping it out of Git and chat replies does not prevent the agent provider from processing it under its own terms.

For manual entry, ask your agent to prepare configuration with placeholders and skip token creation. Create the token in the dashboard, then enter it directly into the gitignored file or CI secret store. Your agent can run the configured build to verify caching without reading or displaying the credential. See tokens and access control for scopes and rotation.

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.

Use the Set up with AI prompt to configure your build tool and verify a real remote cache hit.

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.