Adobe Workfront MCP: turn Linkeddit findings into Workfront work

Adobe's overview says the server connects a Workfront instance to an AI platform such as Claude or ChatGPT, so you can find, create, update and manage items by asking. Its own examples are "Show me all active projects for the Brand Marketing team" and a reminder to approvers who have not responded. The Agent Index lists 68 tools for the Claude connector: 37 reads, 29 writes and 2 it does not classify.

The tools are grouped by area. Approvals has tools such as approvals_get_approval_info and approvals_send_reminder_to_undecided. Planning has planning_search_records, planning_get_record and planning_create_connection_record. Adobe's tools page adds Reporting dashboards, for example reporting_relist_dashboards. Adobe also says the tool list varies by customer, because Planning and Workflow tools depend on your licence.

What it does not have is outside evidence. It reads your Workfront data and nothing about your competitors or your category. That is the gap the Linkeddit connector fills.

Two administrators gate it. Adobe's configure page says a Workfront administrator controls access through two toggles in System Preferences: Read-only MCP tools, enabled by default, and Write MCP tools, disabled by default. If write is off, Claude can still find and read items but cannot create or change them.

On Claude Enterprise, the Claude administrator must also enable the Adobe Workfront connector or allow a custom URL. Your instance must be on Adobe Identity Management System, and each connection ties Claude to a single instance. Adobe says Claude acts with your account, access level and object permissions, so it can do only what you could do yourself.

That default shapes this workflow. With writes off, Claude can still read your open projects and tell you which Linkeddit finding has no task yet, and you create the task by hand. Turn writes on only once you trust the prompt.

Start with a weekly check. Linkeddit's compete_signals_get returns competitor moves with a priority, a category, why it matters and a source link, and compete_monitors_get shows what is watched. On Linkeddit's test account, all five competitor monitors are set to weekly, each with a next run on 2026-11-04. Claude reads the high priority signals and asks Workfront for the open work that already covers them.

Then Claude proposes one task per uncovered signal, naming the competitor, the source link and a suggested owner. A competitive intelligence founder on Reddit described the failure this prevents: "Teams collect screenshots and links, put them into Notion, send a Slack update and move on. Nobody clearly owns the response." Here the response is a Workfront task with an owner and a date.

Keep the output small. Ask for tasks only on signals with high priority, and have Claude list them before it creates anything.

Answer Radar finds the buying questions where an engine recommends a competitor or nobody. Claude reads answer_actions_get, the action queue, which labels each card with a type such as competitor_won or content_gap, an impact estimate and a status. Each card can become one Workfront task, and the card's gap_id goes into the task description so the work stays traceable.

On Linkeddit's test project, plausible.io, the queue returned 2 cards for one question, "Umami vs Matomo for self-hosted analytics". One is competitor_won, "Umami holds the answer", with competitor_citations 12 and samples_absent 3, status in_progress. The other is content_gap, "Not represented", status todo. Each card says what was observed and leaves the fix to the person who owns the page.

When a task is finished, a re-check is a separate step that spends credits. Linkeddit returns an estimate first and Claude waits for your yes before it asks for one.

Linkeddit's gsc_queries_get reads the customer's own Search Console and returns queries with impressions, clicks and average position. Its striking_distance view lists queries at average position 11 to 20 with real impressions, and its ctr_opportunities view lists top five positions that are clicked less than expected. It needs Search Console connected, and says so plainly when it is not.

Claude turns the rows into a short list of pages to refresh, then checks Workfront for an existing request before it proposes a new one, so the same page does not get two tasks. The approval step stays in Workfront: with a page refresh attached to an approval, approvals_send_reminder_to_undecided nudges reviewers who have not answered.

This is a planning aid, not an automation. Claude proposes; a person with write access in Workfront decides what is created.

Start read-only and keep a human in the loop. Adobe's use page says most platforms prompt you before an action runs, especially deletes, and that an ambiguous request can pick the wrong item when two projects have similar names. Ask Claude to name the project by its ID before it writes.

Adobe also says Workfront does not store your prompts or the AI platform's responses, and that whoever provides your AI platform controls how conversations are stored. PromptArmor rates the connector as read and write and lists it for Claude, Claude Code and Claude Desktop. For an enterprise tool that holds approvals, treat writes as a decision your administrator makes on purpose.