pdf2okf·

Wiki

Integrating OKF bundles into agentic tools

The pattern is always the same

An OKF bundle produced by pdf2okf is a directory of plain Markdown files, one per concept, with a metadata index. Every agent on this page can run shell commands and read local files. That overlap is the entire integration: build the bundle once with pdf2okf convert, point any of these agents at the output directory, and they answer from your documents with file-level citations.

No plugin. No API key for the bundle. No custom code.

Hermes Agent (Nous Research, MIT)

Hermes Agent is an open-source autonomous agent built by Nous Research under the MIT license. Its source is at github.com/NousResearch/hermes-agent. It operates entirely via CLI and is designed to execute shell commands, read and write local files, and reason about their contents in an agentic loop.

To use it with an OKF bundle, launch Hermes with a task that includes the bundle path. Hermes searches the concepts directory, reads matching files, and surfaces their paths as citations. No index, no setup: the directory is the interface.

hermes --task "Answer questions from the OKF bundle at ./my-bundle/concepts/ with citations"

Because Hermes Agent is MIT-licensed and runs entirely locally, no document content leaves your machine.

Odysseus (PewDiePie, AGPL)

Odysseus is a self-hosted, local-first AI workspace by Felix Kjellberg (known publicly as PewDiePie), released under the AGPL at github.com/pewdiepie-archdaemon/odysseus. It is built on the premise that models, data, and workspace should all stay on your hardware.

Add the OKF bundle directory as a local knowledge source in Odysseus's workspace config. Queries run against those Markdown files directly. The AGPL license applies only to modifications you distribute; running Odysseus privately on your own hardware carries no open-sourcing obligation.

OpenClaw (Peter Steinberger)

OpenClaw is Peter Steinberger's self-hosted personal AI assistant, published at github.com/openclaw/openclaw. It launched as "Clawdbot", was renamed Moltbot, then renamed again to OpenClaw. It wraps a local model with a shell-accessible interface built around file-based tasks.

Place your OKF bundle directory in a path accessible to OpenClaw and reference it in a query or system prompt. Concept files are plain Markdown; no format conversion is needed.

Claude Code

Claude Code is Anthropic's agentic CLI for working with Claude locally. It runs an agent loop with shell access and file reads, and natively supports MCP servers, which means it will also be able to connect to pdf2okf's planned MCP endpoint once that is released.

Until then, the file path works today:

claude --add-dir ./my-bundle/

Or instruct it inline in a prompt: "Read the files in ./my-bundle/concepts/ and answer with citations." Claude Code treats those Markdown files exactly as it treats any other local knowledge.

Cursor

Cursor is an AI-native code editor with an embedded agent that runs terminal commands and reads workspace files. Add the bundle directory to your Cursor workspace. The agent surfaces concept files via the built-in file search and can cite them directly in its response, with no extra configuration required.

Why self-hosted agents and sovereign bundles belong together

Every agent above (Hermes Agent, Odysseus, OpenClaw, Claude Code, Cursor) is self-hosted, local-first, or both. They share the same core premise as pdf2okf: your tools run on your hardware and your data stays in your control.

Combining a self-hosted agent with an OKF bundle that was produced and stored locally creates a fully sovereign pipeline. The PDF was converted on your machine. The bundle lives on your filesystem. The agent runs locally. No document content, no query, and no answer leaves your infrastructure unless you explicitly choose to route to a remote model for hard multi-hop reasoning via BYOK.

That is the architecture that GDPR-governed environments (legal offices, healthcare providers, public-sector teams) are increasingly asking for. It is also a better engineering default: faster, cheaper, auditable, and offline-capable without any special configuration.

The agent-CLI document pattern, pdf2okf convert then grep, is the simplest version of that architecture. It requires nothing beyond what you already have.

Get started with pdf2okf

Build your first bundle, inspect the concept files, then run the agent you already use against the bundle directory. Join the waitlist at pdf2okf.com to be notified when the CLI is available.

pdf2okf.com

Be there when it opens.

pdf2okf is in private build, self-hosted, sovereign. Leave an email and you'll be first in.