You don't re-train anything. That's the answer, and it's worth being precise about why. A company MCP server built on a knowledge base doesn't bake facts into model weights. The model reads your documents at answer time and cites what it read. So keeping the server current means keeping the documents current — nothing more. On SuperCognit, the knowledge base re-syncs from its sources on a schedule, as often as hourly, so current is something the system maintains rather than something a person remembers to do.
If that sounds too simple, it's because most of the confusion comes from one bad mental model.
Fine-tuning is the wrong frame
People arrive at "our company AI is out of date" carrying a habit from an older workflow: model gets trained, model gets stale, model gets re-trained. That frame makes freshness expensive, so teams ration it. Quarterly updates. A change-request queue for the knowledge base. Someone whose job it is to maintain the AI.
A knowledge-base-backed MCP server works differently. When Claude — or any MCP client — asks your server a question, the server retrieves the relevant passages from your documents and the model answers from those passages, with citations. Change the document and the next answer changes with it. There is no training step to schedule, no weights to touch, no version of the model that "knows" last quarter's prices.
The freshness problem reduces to a syncing problem. And syncing problems are solvable with a schedule.
What actually goes stale
Worth being concrete about, because it determines your cadence:
- Prices and plans — the classic. The website changed; the PDF someone uploaded in March did not.
- Policies — cancellation terms, return windows, service commitments. Legal edits these, and nobody tells the AI team, because there shouldn't need to be one.
- People and structure — who owns what, who to escalate to.
- Product facts — specs, compatibility, what shipped last week.
- Operational details — opening hours, seasonal closures, contact routes.
Notice the pattern: almost all of it already lives somewhere canonical, usually the website. The staleness is a copy problem. The knowledge base was a snapshot, and snapshots age.
Point at the source of truth, then sync to it
Import the live site, not an export
SuperCognit crawls your website into a cited knowledge base in about two minutes. The crawl is the ingestion path, which matters because it can run again — an exported copy can't.
Upload only what has no URL
Internal SOPs, spec sheets, documents that never touch the web. Uploads are the one category that genuinely needs a human to replace them when they change, so keep this pile small and deliberate.
Set the re-sync schedule
Crawled sources refresh on a schedule you choose, as often as hourly. Pages that change often earn a tight cadence; the About page does not.
Let your existing edit flow do the work
From here, updating the MCP server is just updating the website, which your team already does. Marketing edits a page; within the sync window, every connected MCP client answers from the new version.
Choosing a cadence without overthinking it
Hourly sounds like the obvious pick for everything, and for a small site it's fine. But cadence is really a statement about how much staleness you can tolerate per source. Pricing: minutes. Documentation: hours is usually honest. Policy pages: daily catches everything a human legal review actually produces. The goal isn't maximal freshness everywhere; it's that no source's staleness window is longer than the damage a stale answer would do.
The failure mode to design against
The servers that rot are the ones fed by hand. A folder of uploaded PDFs is a knowledge base with no pulse — accurate on upload day, decaying from there, and the decay is invisible until someone gets a confidently cited answer from a superseded document. Citations help here, incidentally: when an answer cites its source, a stale answer at least points at the stale document, so you fix the cause instead of arguing with the symptom.
The discipline, then: anything that changes on the web gets synced from the web. Anything uploaded gets an owner and a review habit. Nothing gets pasted in as free text that has a canonical home elsewhere.
One build, every surface
A detail that makes the economics work: the same build produces both a chat agent and an MCP server. One knowledge base, one sync schedule — the answer a visitor gets on your site matches the answer a colleague gets in Claude. You maintain freshness once and every surface inherits it. Which is the whole point: how current your company MCP is should be a property of the plumbing, not a task on somebody's list.