zim-mcp is an open source MCP server that exposes Kiwix ZIM archives, including offline Wikipedia, to an AI agent over the Model Context Protocol. You download a ZIM file, run the server locally, and your agent searches and reads it with no API key, no internet connection and no per-call fee. It is a local knowledge source an agent can ground answers in.
zim-mcp MCP server (offline Wikipedia for AI agents)
ZIM archive (Wikipedia, Kiwix download)
|
v
+------------------------------+
| local MCP server (zim-mcp) |
| search + read over MCP |
+------------------------------+
| stdio or HTTP, on your machine
v
AI app: Claude, Cursor, Codex, VS Code
|
v
grounded answer, no API key, no internetWhat Is the zim-mcp MCP Server?
zim-mcp is a local MCP server that reads Kiwix ZIM archives and serves their articles to an AI agent. It is open source and published at github.com/gordonbrander/zim-mcp.
Kiwix packages Wikipedia and many other reference works into single ZIM files that you can read with no connection. zim-mcp puts those files behind the MCP protocol. Your agent asks for a search or an article the same way it calls any other MCP tool, and the server answers from the file on disk.
The important details, all from the project: it serves local archives, speaks MCP, and stays on your machine. There is no account, no hosted endpoint and no key to rotate. That is a different shape from a hosted search API, where every lookup is a network call you pay for.
How zim-mcp Serves Offline Wikipedia to Your AI Agent
The flow is three steps: get an archive, start the server, point your agent at it. Nothing in the middle touches the network.
- Download a ZIM file. Kiwix offers Wikipedia and other reference works as downloadable archives. Save one to a folder on the machine where your agent runs.
- Start zim-mcp against that file. The server opens the archive and registers its search and read operations as MCP tools.
- Connect your AI app. Claude, Cursor, Codex, Windsurf or VS Code connects to the server the same way it connects to any other MCP server, and the Wikipedia tools appear in its tool list.
After that, an agent asked “what year did the Treaty of Waitangi get signed and who signed it” can call the local tool, read the article, and cite it. The answer comes from the archive you hold, not from a model’s memory, which is where most confident wrong facts come from.
Can You Run zim-mcp on Your Own Laptop or Server?
Yes. zim-mcp is a local process, so it runs on the same laptop, desktop or server that runs your agent. The only real cost is disk space for the archive.
That matters for two reasons. First, a laptop with no signal on a train can still ground an agent in an encyclopedia. Second, nothing leaves the machine, so private research and client work stay private.
A practical setup keeps the archive on fast local storage, restarts the server with the agent, and treats it as one more entry in the agent’s MCP config next to any hosted servers you use. Creator OS’s own hosted server slots into the same configs, described at https://www.creatoros.ca/docs/mcp.
Why Free, Private Knowledge Sources Matter for Social Media AI
An agent that posts for you has two knowledge problems: it needs facts, and it needs to check the facts it already wrote. Offline reference material solves both without adding a paid dependency.
Most agent stacks today get their grounding from a hosted search API. That works, and it is fast. It also means every fact check is a metered call, a data-sharing relationship, and a failure mode when the network or the vendor is down. A local ZIM archive is the other end of that trade: slower to update, free to query, impossible to rate limit.
For a creator running accounts on 11 platforms plus Skool, that difference compounds. A reply with a wrong date or a wrong attribution is a public mistake on a client’s page. The cheaper and more private you can make the checking step, the more often the agent will actually do it.
zim-mcp for Fact-Checking Replies and Drafting Accurate Posts
Give the agent a rule: before it publishes a claim about a person, place, date or event, it checks the local archive. Then wire that rule into your actual posting flow.
Here is what it looks like in practice with the Creator OS CLI. The agent drafts a caption, checks the claims against zim-mcp, then publishes:
creatoros validate:post-length --text "In 1887 ... "
creatoros posts:create --text "<caption>" \
--platforms instagram,tiktok,youtube \
--media https://cdn.example.com/clip.mp4 \
--scheduledAt 2026-03-04T17:00:00Z --timezone America/Toronto
The same habit applies to replies. When a comment on your post makes a factual correction or asks a question about a historical claim, the agent reads the thread, checks the local archive, and answers from it:
creatoros inbox:post-comments post_123 --accountId acc_456
creatoros inbox:reply post_123 --accountId acc_456 \
--commentId cmt_789 --message "<verified answer>"
Reply support in Creator OS covers Instagram, Facebook, X, Threads, YouTube, LinkedIn and Reddit. Comment to DM funnels run on Instagram and Facebook. The full inbox flow is at https://www.creatoros.ca/docs/use-cases/comments-to-dm.
For a longer written piece, the same check precedes a blog draft. Connect a WordPress site or a Shopify store blog and the AI writes the article, drafts stay as drafts unless you set publish, and the claims were grounded before the first sentence was written. See https://www.creatoros.ca/docs/use-cases/seo-blog.
How zim-mcp Cuts Per-Call Costs for Agents That Run Your Accounts
Every hosted lookup an agent makes has a price. A local archive has none, so the checking step stops competing with the posting step for budget.
Think about a week of agent work on a busy account: fifty replies drafted, ten captions fact-checked, a handful of blog sections verified. Against a metered search API, every one of those is a line item. Against a local ZIM file, the marginal cost is zero, which changes the agent’s behavior. You can tell it to check every factual claim, and it can.
The same logic shows up in the way agents spend tokens. Work that never leaves your machine costs nothing to repeat, so a prompt like “check every claim in this thread against the local archive before you reply” is cheap enough to make the default. Keep the metered, open-web lookups for the few tasks that need this week’s news, and move stable reference checks to the archive you already hold. Our earlier post on cutting AI agent research tokens covers the same trade in a different place.
A simple division of labor works well:
| Task | Local zim-mcp archive | Hosted search API |
|---|---|---|
| Stable facts (history, geography, definitions) | Good fit | Works, metered |
| News and prices from this week | Not the right tool | Needed |
| Cost per lookup | None after download | Per call |
| Runs without a connection | Yes | No |
| Data leaves your machine | No | Yes |
zim-mcp vs Hosted Search APIs
These are different tools for different jobs, and most useful agent stacks end up with both. A hosted search API answers questions about the live web. zim-mcp answers questions about stable reference material, from a file you control.
When you choose what to install, ask three questions: does this need to be fresh, does it need to be private, and does it need to be cheap to call a hundred times a day. A knowledge source that fails the freshness test can still be the right home for the other ninety percent of checks. The wider question of what shipping a server means for agents is covered in shipping an iOS app with the MCP App Store Connect server.
Creator OS ships a hosted MCP server and a CLI so an AI agent can run your accounts. The agent keeps its own tools; Creator OS handles publishing, replies, DMs, analytics and Skool, and the two connect over the same protocol. That means a local Wikipedia server and a hosted publishing server sit in one config, and the agent decides which to call.
FAQ
What is zim-mcp and how does it work?
zim-mcp is an open source MCP server that exposes Kiwix ZIM archives, including offline Wikipedia, to an AI agent. You download an archive, run the server locally, and your agent calls its search and read tools like any other MCP tool. The project lives at github.com/gordonbrander/zim-mcp.
Does zim-mcp need an internet connection or API key?
No. Once the ZIM archive is on disk, the server answers from that file, so no connection and no API key are involved. That is the whole point of the setup: private, free, local grounding for an agent.
How do I install the zim-mcp MCP server with Claude or Cursor?
You add it as an MCP server in your AI app’s configuration, pointed at the archive on your machine, and the same pattern applies in Cursor and VS Code. The exact install steps belong to the project’s own README at github.com/gordonbrander/zim-mcp. For the Creator OS side of a config, the docs at https://www.creatoros.ca/docs/mcp cover Claude, ChatGPT, Claude Code, Cursor, Windsurf and VS Code.
Can an AI social agent use offline Wikipedia for fact-checking?
Yes, if you give it the rule and the tool. The agent checks a claim against the local archive before it calls the publish or reply step. In Creator OS that means checking first, then running creatoros posts:create or creatoros inbox:reply with the verified text.
zim-mcp vs Higgsfield AI MCP server: which should I install?
They do different jobs. zim-mcp is a knowledge source: offline reference material an agent reads for grounding. Higgsfield AI’s MCP server is described on its own site, so read that page for what it does. Judge each one on whether your agent needs reference facts or a creative capability, and install whichever fits that gap.
Get Started With Creator OS
Creator OS gives an agent the publishing side: posts, replies, DMs, analytics, Skool and ads, reached through a hosted MCP server, a CLI, a REST API or the web and iOS apps. Pair it with a local knowledge server like zim-mcp, and the agent has both the facts and the hands.
Plans start at $14.99/month for 11 platforms plus Skool, with the API, the MCP server and the CLI included. Sign up at https://www.creatoros.ca/sign-up, and watch the walkthroughs on the KevBuildsApps YouTube channel.
Get started for $0
Free trial for new subscribers. Cancel anytime.