Harness Wiki

Skill-first tooling to scaffold and maintain an LLM Wiki — a long-lived, agent-maintained markdown knowledge vault.

You curate sources and ask questions; the agent summarizes, cross-references, files, flags contradictions, and keeps an index and log current. Unlike file-upload RAG, an LLM Wiki compiles knowledge into durable, interlinked pages instead of retrieving chunks from scratch on every query.

Based on Andrej Karpathy's LLM Wiki concept — see Credits.

Companion product. @codemeall/harness-bridge tracks the current in-flight task in one repo via .harness/bridge.md. Harness Wiki tracks the durable knowledge around many tasks — research, design rationale, decisions. Use them together: wiki for the why, project repos for the what, bridge files for the current work.

Quickstart

In any agent that can load a skill (Claude Code, Codex, …), open the folder you want to become your knowledge base and say:

Use the Harness Wiki skill to initialize an LLM Wiki vault here. Name: knowledge. Domain: <your topic>.

Then just work with it in plain language:

1. Feed it — drop a file into raw/, then say ingest raw/<file> (or ingest <url>).
2. Ask it — ask any question; the agent answers from the wiki with citations.
3. Maintain it — say lint for a periodic health check.

That's the whole loop. You curate and ask; the agent does the summarizing, cross-referencing, and filing. Want to see the output before installing anything? Browse examples/knowledge-vault/ — a small, complete vault.

The wedge: wiki ↔ repo ↔ bridge

Notes apps store text; Harness Wiki stores decisions and the work they feed. A vault registers the sibling project repos it serves (vault_link), so design rationale connects to implementation: the wiki holds the why, the project repo holds the what, and harness-bridge holds the current work. When a synthesis is approved, it hands off to the paired repo for implementation — and substantive decisions made there flow back into the vault. vault_doctor keeps the structure honest: stale source pages (the raw file changed since you ingested it) and broken project links get flagged on every health check.

What you get

harness-wiki gives agents a portable skill for the full vault lifecycle. MCP and CLI commands are optional accelerators for deterministic scaffold/status operations; the actual ingest/query/lint operations are agent-driven, guided by the skill and the schema written into CLAUDE.md / AGENTS.md.

raw/                       # immutable source documents you curate (assets in raw/assets/)
wiki/entities/             # people, orgs, products, places — one page each
wiki/concepts/             # ideas, frameworks, methods, patterns
wiki/sources/              # one summary page per ingested raw source
wiki/syntheses/            # comparisons, analyses, decisions worth keeping
wiki/index.md              # catalog of every page
wiki/log.md                # append-only operations log
CLAUDE.md / AGENTS.md      # the agent schema (carries the vault marker)

Installation

Install the packaged skill into your agent's skills directory with one command — no global install, no manual file copying:

npx -y @codemeall/harness-wiki skill-install claude          # Claude Code (~/.claude/skills)
npx -y @codemeall/harness-wiki skill-install claude-project  # this project's ./.claude/skills
npx -y @codemeall/harness-wiki skill-install codex           # Codex (~/.codex/skills)
npx -y @codemeall/harness-wiki skill-install --target <dir>  # any custom skills directory

| Preset | Installs to | |---|---| | claude | ~/.claude/skills (Claude Code, personal) | | claude-project | ./.claude/skills (Claude Code, this project) | | codex | ~/.codex/skills (Codex, personal) | | agents | ~/.agents/skills (generic agents, personal) | | local | ./.agents/skills (generic agents, this project) |

Once installed, open the folder you want to become a vault and tell your agent: "Use the Harness Wiki skill to initialize an LLM Wiki vault here."

For MCP-capable agents, also run the server (no install):

npx -y @codemeall/harness-wiki mcp

For direct terminal setup without installing globally:

npx -y @codemeall/harness-wiki vault-init --name "knowledge" --domain "software project"
npx -y @codemeall/harness-wiki vault-status

Or install globally for the CLI:

npm install -g @codemeall/harness-wiki

From a Git checkout:

git clone https://github.com/codemeall/harness-wiki.git
cd harness-wiki
npm install
npm run build

Requires Node.js >= 20.

Usage

Skill-first usage

Load skills/harness-wiki/SKILL.md into your agent or harness, then open the directory you want to become a vault and say:

Use the Harness Wiki skill to initialize an LLM Wiki vault here.
Name: "knowledge"
Domain: software project

The skill covers initialization, ingest, query, lint, paired projects, and schema maintenance. Once the vault exists, the generated CLAUDE.md / AGENTS.md schema becomes the local operating contract.

Codex-style skills: install or copy skills/harness-wiki/SKILL.md into your skills directory, then ask Codex to use the Harness Wiki skill in the vault directory.

You can install the packaged skill with the CLI:

npx -y @codemeall/harness-wiki skill-install claude
npx -y @codemeall/harness-wiki skill-install claude-project
npx -y @codemeall/harness-wiki skill-install codex
npx -y @codemeall/harness-wiki skill-install agents
npx -y @codemeall/harness-wiki skill-install local

Use --target <skillsDir> for a custom harness skills directory:

npx -y @codemeall/harness-wiki skill-install --target ./.agents/skills

Claude-style project instructions: add the skill text to project instructions or paste it at the start of the session. After initialization, Claude should follow the generated CLAUDE.md.

ChatGPT/custom harnesses: load the skill as the system/developer instruction or first user message for the session. If your harness supports local tools, expose MCP as described below.

Generic agents: paste the skill text, then point the agent at the target directory. The skill includes the manual non-MCP setup path.

MCP accelerator

Register Harness Wiki as an MCP server when your agent supports MCP. The skill will use MCP for deterministic scaffold/status operations. Example MCP client config:

{
  "mcpServers": {
    "harness-wiki": {
      "command": "npx",
      "args": ["-y", "@codemeall/harness-wiki", "mcp"]
    }
  }
}

From a local build, point at the built CLI instead:

{
  "mcpServers": {
    "harness-wiki": {
      "command": "node",
      "args": ["/absolute/path/to/harness-wiki/dist/cli.js", "mcp"]
    }
  }
}

Then open the directory you want to become a vault and tell the agent:

Use the Harness Wiki skill to initialize an LLM Wiki vault here.
Name: "knowledge"
Domain: software project

When calling MCP tools, the agent should pass the target workspace absolute path as cwd. This avoids initializing the MCP server launch directory instead of the project directory.

MCP tools

| Tool | Purpose | |---|---| | vault_init | Scaffold the vault at optional cwd: creates raw/, wiki/{entities,concepts,sources,syntheses}/, CLAUDE.md, AGENTS.md, wiki/index.md, wiki/log.md. Auto-detects empty dir (standalone vault) vs. existing project (appends a marker-fenced <!-- harness-wiki:vault-schema --> block to the host CLAUDE.md / AGENTS.md without touching other content). Idempotent on re-run via the marker. | | vault_status | Report scaffold state at optional cwd and list files in raw/ with no matching wiki/sources/<slug>.md (pending ingests). Filename/slug matching is normalized, so PRD Checkout V2.pdf pairs with prd-checkout-v2.md. | | vault_doctor | Deterministic structural health check: pages missing YAML frontmatter, dead [[wikilinks]], orphan pages with no inbound links, and pages absent from wiki/index.md. The mechanical first pass of a lint. | | vault_search | Plain-text search across all wiki/ pages (case-insensitive), returning file + line + matching text. For locating relevant pages in a large vault before answering a query. | | vault_stamp | Record source_file + source_sha256 into source-page frontmatter so vault_doctor can detect when a raw source file changes (a stale page). Run after each ingest; re-run to acknowledge a re-review. | | vault_link | Register a sibling project in the Paired-projects table and write a marker-fenced back-pointer into that repo's CLAUDE.md. The wiki↔repo wedge — discoverable from both sides, idempotent. |

MCP prompt: wiki_init_prompt returns the canonical setup/operations prompt (also available as the resource harness-wiki://prompts/vault-init.md).

CLI

harness-wiki mcp                                              # run the MCP server (stdio)
harness-wiki vault-init --name "knowledge" --domain "software project" [--cwd <path>] [--force]
harness-wiki vault-status [--cwd <path>]
harness-wiki vault-doctor [--cwd <path>]                      # structural health check
harness-wiki vault-search "<query>" [--cwd <path>] [--limit <n>]
harness-wiki vault-stamp [--cwd <path>]                       # record source hashes for staleness
harness-wiki vault-link --name <n> --path <p> --purpose <text> --status <text> [--cwd <path>]
harness-wiki skill-install <codex|agents|local> [--force]
harness-wiki skill-install --target <skillsDir> [--force]

From a Git checkout, substitute node /absolute/path/to/harness-wiki/dist/cli.js for harness-wiki.

harness-wiki mcp is a stdio server command. It looks idle when run directly because it is waiting for an MCP client; configure it in your agent instead of using it as an interactive command.

Prompt fallback

If your agent cannot load skills, paste prompts/vault-init.md as the first message in the directory that should become the vault, and answer the agent's domain questions.

Documentation

• docs/llm-wiki.md — full setup, daily ingest/query/lint workflow, and when to use the wiki vs. a project repo.
• docs/example-feature-workflow.md — a worked flow: tracking features, a pending list, moving items to done, reopening them, and asking about a feature's options.

Credits

The LLM Wiki pattern this tool implements was described by Andrej Karpathy in his LLM Wiki gist — an idea file meant to be handed to an LLM agent, which then builds out the specifics with you. Harness Wiki packages that pattern as a portable skill, MCP server, and CLI so it's repeatable across agents and projects. Credit for the underlying concept is his; this repository is an independent implementation.

License

MIT