Proposal automation

Upwork has no single rule called proposal automation. It has rules for each way software can touch a proposal, and the clearest statement is the MCP article: an AI agent can help draft a proposal, and before submission "you'll review the proposal details and explicitly confirm". Drafting is allowed; submitting stays a confirmed human act on that route.

Steps that never touch Upwork. Scoring a job, choosing which member fits, drafting the letter and the screening answers, and deciding a bid can all run on your own systems from data you hold. Upwork's automation rules are about activity on Upwork, so these steps fall outside them. They still carry a quality risk: a client can flag a cover letter as spam.

Steps that act on Upwork. Searching for jobs, reading posts and client histories, and submitting the form. For these Upwork names two permitted routes: an approved API key used within its approved use case, and its MCP server with a confirmation on every write. Scripts, extensions and hosted services acting on a signed-in session are what its automation article calls bots.

A rule worth borrowing. When an MCP proposal confirmation fails, Upwork says not to assume it was submitted and not to retry repeatedly, but to check submitted proposals and the Connects balance first. Any automated send should behave the same way, because a duplicate proposal costs Connects and looks like spam.

How Upsight uses it. Upsight automates matching, scoring, drafting and sending per campaign, and files proposals through Upwork's official API with the workspace's own key. A campaign can require a person to approve every proposal, and test mode runs the whole path while filing nothing.