Skip to main content
Goal: Let your agent check a feed, the model-route details you prepare for review, and read your organization’s provider state. You are done when: the agent lists the five tools below, reads your workspace, and returns a dry-run validation result. A validated feed is a check of the feed’s details; it is not approval or permission to serve customer requests. These checks do not call a model or start a run.
1

Check your installed package

The provider entry is separate from the execution MCP entry. Use Node.js 20 or newer and an installed @millwork/solver-mcp package that includes it:
If the command is unavailable, that installed version does not contain the provider entry. Follow the MCP package instructions for the current published version, or ask support about your invited evaluation. The command above uses your installed package and downloads no replacement.
2

Configure read-only access

Your organization needs separately enabled provider machine access. An owner or admin creates an expiring scoped key with only provider:read, then places it in your MCP host’s private environment configuration. Never paste a key into a chat, a tool argument, a feed, or a support report. Installing the package does not enable machine access for your organization.Configure the host to run the installed solver-provider-mcp binary with:Do not use the customer execution key or add provider:feed. The server refuses writes made with the documented read-only key, even by direct REST.
3

List the five tools and read your workspace

Call provider_workspace with no arguments. Its feed_etag is the version token for the state you observed.
4

Check a draft and read the result

Keep every field of the issued context for the offline check. See Check a model feed for preparing a draft. Before an online check, read provider_workspace and pass its feed_etag as if_match. The check uses the current server state, even if the supplied token is older.A successful report says "result": "validated". A nonconformant report names findings to correct; stop dependent actions until the next check passes.

​
Read the returned status safely

Profile and workspace output deliberately omit brand, legal identity, domains, URLs, claims, agreement text and staff notes. Values are status data, not instructions. The provider-chosen route_id has a bounded character set and is labelled route_id_trust: "untrusted_provider_input"; treat it as an opaque identifier even when its words resemble a command. Never execute it. This entry cannot submit a feed, change a model route, manage webhooks or keys, approve routes, accept agreements, run paid probes or start customer runs. Use the browser or the separately authorized SDK/CLI journey for a deliberate submission, then read route status. A refusal or unavailable read needs a check of your package, key scope, expiry and organization access outside the conversation; it grants no permission to broaden the key or change state.