English | 中文
An agent-first, local-first, terminal-native harness for maintaining OKF-compatible LLM Wikis.
OKF Harness is an independent open-source project built on two upstream ideas: Andrej Karpathy's LLM Wiki pattern for agent-maintained living wikis, and Google's Open Knowledge Format / OKF specification for portable markdown knowledge bundles.
source files or URLs
|
v
raw/sources + .okfh/manifest.jsonl
|
v
wiki/*.md with citations
|
v
supported agents use okfh evidence/read/graph
OKF Harness does not ask you to learn a new knowledge-base app. You run setup once for the agents you use, create one local workspace per knowledge domain, then ask your agent to add sources, maintain the wiki, and answer from it.
OKF Harness builds on:
- Andrej Karpathy's LLM Wiki: the agent-maintained wiki pattern of index, log, linked pages, ingest, query, and lint.
- Google's Open Knowledge Format announcement and OKF specification: the markdown-plus-frontmatter bundle shape that keeps knowledge portable across tools.
This repository is not affiliated with or endorsed by Andrej Karpathy or Google.
If you're on macOS or Linux, run this script:
curl -fsSL https://okf-harness.dev/install.sh | shIn Windows PowerShell, run this instead:
irm https://okf-harness.dev/install.ps1 | iexAlready have Node.js 22 or newer?
npx @okf-harness/setup@latestNormal use needs Node.js 22 or newer; git; the shared okfh runtime; and at least one supported agent integration. Repository development additionally needs pnpm.
Setup installs or updates the shared global okfh runtime after confirmation, detects supported agent clients, and installs the selected native integrations. Direct native install paths are available for users who already know their agent:
| Agent | Native install command |
|---|---|
| Claude Code | claude plugin marketplace add pumblus/okf-harness && claude plugin install okf-harness@okf-harness |
| Codex | codex plugin marketplace add pumblus/okf-harness --json && codex plugin add okf-harness@okf-harness --json |
| OpenCode | opencode plugin @pumblus/okf-harness --global |
| Pi | pi install npm:@pumblus/okf-harness |
| Hermes Agent | hermes skills tap add pumblus/okf-harness && hermes skills install pumblus/okf-harness/okf-harness |
| OpenClaw | openclaw skills install @pumblus/okf-harness --global |
Advanced direct CLI runtime installation is documented in the CLI reference. It does not write agent bootstrap entrypoints.
After setup, start from the global bootstrap entrypoint in your current agent. If that agent has workspace-local guidance, the bootstrap handoff tells you how to refresh into the workspace-local okf-harness entrypoint.
The recommended parent folder is only a convention, not a hidden CLI default. On macOS or Linux, use $HOME/Documents/OKF Harness. On Windows PowerShell, use $env:USERPROFILE\Documents\OKF Harness. On Command Prompt, use %USERPROFILE%\Documents\OKF Harness.
Use the OKF Harness entrypoint name exposed by the agent you already use. The entrypoint name is stable; the calling syntax belongs to the agent. Codex usually uses $okf-harness, Claude Code usually uses /okf-harness, and other native integrations expose the same entrypoint name through their own skill or plugin UI.
When copying a prompt below, replace the bracketed entrypoint with your agent's actual invocation.
<okf-harness-bootstrap> Set up a workspace for my AI research notes in my Documents folder, then tell me how to refresh this agent context.
After setup or workspace selection, work inside the workspace with the workspace-local entrypoint:
<okf-harness> Check this workspace and tell me whether it is ready.
Bootstrap can also discover or select a workspace from a local workspace collection and repair current-agent setup for the selected workspace. It does not synthesize wiki content, migrate non-empty non-workspace folders, or write global root guidance files.
For a transient diagnostic command:
npx --package @okf-harness/cli okfh doctor --jsonThis does not add a global okfh binary.
Add a source:
<okf-harness> Add this PDF to my workspace, update the wiki with citations, then check the workspace again.
Ask a question:
<okf-harness> What does my workspace say about LLM Wiki structure?
Most personal knowledge tools make the app the center. OKF Harness makes the local folder the center:
- raw source material stays inspectable under
raw/sources/ - synthesized knowledge lives in ordinary markdown under
wiki/ - citations connect topic pages back to reference pages and source IDs
okfh --jsongives agents a deterministic tool surface- the graph report is a local HTML file, not a hosted service
The recommended layout is one workspace per knowledge domain, research area, or privacy boundary. Keep them under a local Documents/OKF Harness/ folder unless you have a reason to separate them.
The product stays narrow on purpose: local files, terminal-native commands, bounded evidence, bounded reads, and explicit provenance come first. Broader surfaces such as GUI, cloud sync, Obsidian helpers, source connectors, and vector retrieval belong in the roadmap only when they preserve those guarantees.
- Initializes a local OKF Harness workspace.
- Installs supported workspace guidance for agents with workspace adapters.
- Registers files and URL pointers as raw sources.
- Produces ingest plans so an agent can update the wiki with citations.
- Prepares bounded evidence briefs from synthesized wiki pages before answers.
- Searches and reads synthesized wiki pages for debugging and bounded continuation.
- Checks OKF conformance and Harness lint findings.
- Generates a self-contained graph report.
The agent uses okfh --json through your local shell. For example:
- first setup uses the global bootstrap entrypoint to resolve the workspace collection, confirm writes, call
okfh initwith the current agent adapter, and return agent context refresh guidance - ingest calls
okfh source addandokfh ingest plan - answers use
okfh evidence, then at most one boundedokfh readwhen a continuation cue is needed - validation uses
okfh check - graph reports use
okfh graph
Developers can call the CLI directly when they need to script or debug a workspace:
okfh check --workspace "$HOME/Documents/OKF Harness/ai-research" --json
okfh evidence "LLM Wiki" --workspace "$HOME/Documents/OKF Harness/ai-research" --json
okfh search "LLM Wiki" --workspace "$HOME/Documents/OKF Harness/ai-research" --json
okfh read topics/llm-wiki --workspace "$HOME/Documents/OKF Harness/ai-research" --json
okfh graph --workspace "$HOME/Documents/OKF Harness/ai-research" --jsonIf the okf-harness-bootstrap entrypoint is missing, stale, or blocked by an unmanaged same-name skill, run:
okfh doctor --jsondoctor reports runtime, native integration, legacy bootstrap fallback, and workspace checks separately. Use okfh bootstrap status|repair --agents codex|claude|all --json only as advanced legacy fallback repair tooling for Claude/Codex adapters; the primary first-setup workflow is setup plus the agent prompt above.
- Workflows explains the user-facing agent flows, including the first-start check.
- CLI reference lists commands, options, and JSON behavior.
- Roadmap shows the current focus and demand-ranked ideas.
- Example workspace gives a small lintable workspace.
- Contributing explains project scope and verification.
- Security explains local data boundaries and reporting.
pnpm install
pnpm docs:llms
pnpm test
pnpm typecheck
pnpm buildSee CONTEXT.md for the project glossary and docs/adr for architecture decisions.
Thanks to Andrej Karpathy for publishing the LLM Wiki pattern, and to Google for publishing Open Knowledge Format as a simple, portable shape for markdown knowledge bundles. OKF Harness adapts those ideas for a local, agent-first workflow.
Thanks also to Tw93's Waza and Matt Pocock's Skills for Real Engineers for shaping the development behind this project.
Apache-2.0. See LICENSE.