OKF for Gemini — Using Knowledge Bundles with Google Gemini

What Is OKF?

OKF, short for Open Knowledge Format, is an open standard for packaging knowledge as structured, human-readable markdown. It was launched by Google Cloud in mid-2026 and was designed from the start to be model-agnostic. An OKF bundle is simply a directory of plain-text documents with three defining traits:

• YAML frontmatter on every document, declaring its title, description, tags, and version
• An index.md entry point that acts as the bundle's manifest and table of contents
• Cross-links between concept documents, forming a navigable knowledge graph

Because everything is plain text and markdown, any large language model — including Google's Gemini — can read an OKF bundle with zero special tooling. OKF is not tied to a single vendor, which is exactly why it has spread quickly across the AI agent ecosystem. It works just as well in Claude Code, OpenCode, Cursor, and other coding agents as it does with Gemini.

For a deeper look at the format itself, including its history and structure, see our guide to Open Knowledge Format.

Why Gemini Users Need Knowledge Bundles

Gemini is extraordinarily capable at reasoning, but it is not omniscient. When you ask it about your codebase, your domain, your team's conventions, or a niche subject, it has to guess — or you have to paste in context. That is where knowledge bundles come in. An OKF bundle gives Gemini a curated, structured source of truth it can consult instead of inventing answers.

Consistent, Reusable Context

Most people give Gemini context by copy-pasting chat history or re-explaining themselves every session. That is slow, error-prone, and non-repeatable. An OKF bundle packages that context once and in a versioned, searchable way. You point Gemini at the bundle and it immediately understands the domain, conventions, and constraints without you retyping anything.

Structured Knowledge Over Raw Text

The structure matters as much as the content. Because OKF documents have frontmatter, headings, and cross-links, Gemini can navigate the bundle like a book rather than reading a wall of text. It knows what each document is about before it reads it, and it can jump between related concepts. That structured retrieval produces more accurate, better-grounded answers than dumping files into the chat.

Team Knowledge That Survives

Bundles are stored in Git and shareable across a team. When a teammate updates a bundle, everyone gets the updated knowledge. This turns tribal knowledge — the stuff living in one person's head — into documentation Gemini (and other agents) can actually use.

OKF and the Google Gemini Ecosystem

OKF and Gemini share a common origin: both come from Google. OKF was launched by Google Cloud, and Gemini is Google's flagship model family. That relationship matters for a few practical reasons:

• Format alignment — Gemini natively handles the markdown that makes up OKF bundles, so no conversion is required.
• Tooling direction — Google's agent tooling, including its work on the Model Context Protocol, treats OKF as a first-class way to inject knowledge.
• Cross-vendor portability — A bundle created for Gemini can be handed to a Claude, OpenCode, or Cursor agent later without any rewrite. The standard is open.

If you are deciding how to structure knowledge for a Gemini-based workflow, OKF is the natural choice: it is the format Google backs, it is open, and it plays nicely with the broader agent ecosystem already indexed on BundleDex.

For more on how OKF differs from alternatives like retrieval-augmented generation, see our OKF vs RAG comparison.

How to Load and Use OKF Bundles with Gemini

There are three practical ways to get OKF knowledge into Gemini. Which one you choose depends on where you use Gemini — in the browser, through its API, or inside a connected agent.

1. Load via the MCP Endpoint

The most powerful option for agent workflows is the Model Context Protocol (MCP). BundleDex exposes an MCP server at /api/mcp that lets Gemini-connected agents fetch bundle content on demand, right when they need it, instead of you pre-loading everything. Adding it to a Gemini agent that supports MCP is straightforward:

# Add the BundleDex MCP endpoint to your agent config
mcp-add https://bundledex.net/api/mcp

# In clients that use a JSON server config, register:
#   "mcpServers": { "bundledex": { "url": "https://bundledex.net/api/mcp" } }

Once connected, you can ask Gemini to retrieve a bundle by name or slug, and it will pull the relevant documents from the MCP server and reason over them in context.

2. Upload Bundle Files Directly

For one-off sessions in the Gemini web or mobile app, the simplest path is uploading the bundle's markdown files directly:

# Clone a bundle from its repository
git clone https://github.com/iwe-org/iwe.git ~/okf-bundles/iwe

# Then drag the folder / markdown files into the Gemini chat.
# Gemini reads index.md and the concept documents natively.
# Tips:
#  - Upload the whole folder, not just one file
#  - Ask Gemini to read index.md first so it maps the structure

Gemini handles markdown natively, so this works with any OKF-conformant bundle — no conversion needed. For the best results, upload the entire bundle directory rather than individual files so Gemini can follow the cross-links.

3. Pass Context Through the Gemini API

If you are building your own application on the Gemini API, you can read a bundle's documents in your application code and inject them into the prompt or system instructions:

# Pull the bundle's docs into your app and assemble context
# (pseudo-code)
const bundle = await readOkfBundle("~/okf-bundles/iwe");
const systemPrompt = bundle.index.md + "

" + bundle.documents.join("

");

# Send to the Gemini API with that context prepended

This gives you fine-grained control over what Gemini sees, how much context you spend, and how the bundle is updated over time.

Best Practices

• Start with index.md — Tell Gemini to read the manifest first so it understands the bundle's structure before diving into details.
• Prefer conformance — Look for the OKF badge on BundleDex to ensure bundles follow the full spec and validate cleanly.
• Combine bundles — Layer a conventions bundle, a domain bundle, and a project bundle to cover different kinds of context in one workflow.

OKF vs Gemini Context Windows

Gemini models ship with large context windows, and it is tempting to treat that as license to paste in everything. But a big context window is not a substitute for good knowledge structure. Here is what OKF gets you that raw context does not:

Token Efficiency

A packed chat or a pile of pasted files burns tokens fast. OKF bundles are concise, well-scoped markdown. When you load a bundle through an agent, you can retrieve only the relevant concept documents rather than dragging in an entire codebase or a huge chat history. That keeps token spend down and leaves more room in the context window for the actual task.

Searchable, Not Serial

Raw context is serial — the model reads it start to finish. An OKF bundle is navigable. Gemini can look up the specific concept it needs, follow a cross-link to a related document, and assemble an answer without wading through everything. This matters even more as context windows fill up: structured knowledge degrades gracefully; unstructured dumps degrade into confusion.

Retrieval vs Bundles

If you want a full comparison of embedding-based retrieval and structured bundles, see our OKF vs RAG guide. The short version: RAG retrieves snippets at query time, while OKF bundles are curated in advance. They are complementary — and the MCP endpoint at /api/mcp is a clean way to layer retrieval over bundle content.

Practical Sizing

There is no need to stuff your entire company wiki into one bundle. Keep bundles focused on a domain, and let the context window hold the bundle plus the active task. A well-scoped bundle of a few dozen documents fits comfortably in Gemini's window and leaves headroom for the reasoning that actually drives the work.

Best OKF Bundles for Gemini

These bundles are tagged with Gemini / Google relevance or mention Gemini in their description. They range from general knowledge systems to Gemini-specific workflows:

Browse all 583 bundles on BundleDex, or connect Gemini to the BundleDex MCP server at /api/mcp to pull bundle content on demand.

FAQ