Show HN: A Karpathy-style LLM wiki your agents maintain (Markdown and Git)

Who Should Read

A developer running several Claude Code or Codex agents who finds per-agent approvals and cost tracking tedious

Core Mechanics

• Title and README don't match
  • The title says a Karpathy-style LLM wiki on Markdown and Git.
  • The README actually describes gawkbot, an agent management tool.
  • The README never explains how a wiki is maintained.
• One sentence builds a bot
  • Describe a job in a sentence or demo it once, and the AI builds the bot.
  • Each bot gets its own screen, schedule and tools.
• Every external write needs human approval
  • Reads are free; email, Slack, CRM writes, commits, purchases and deletes wait for approval.
  • After approval it runs 24x7.
• It finds agents already on your machine
  • It checks PATH, config folders and the process list for Claude Code, Codex, Opencode, Ollama and others.
  • It says it launches nothing while looking.
  • Adopted agents run on their own sign-in, with gawkbot's keys kept out of their environment.
• Memory is local files
  • Bot knowledge, run transcripts and the company brain live as files under ~/.wuphf/.
  • Knowledge pages cite sources, and no backend or API key is needed.
• The README states cost and status openly
  • A typical bot build costs $1–2 on Claude Code.
  • A table marks each feature shipped, partial or planned.
  • The iPhone app is written but not yet built in CI.

Community Reactions

• Title doesn't match the repo, one commenter said
  • A commenter pointed out that the post title doesn't match the repo at all.
  • Expecting a Karpathy wiki gets you something different.
• Worry that agent-written notes rot
  • One commenter said entries agents wrote without human review became confidently wrong after six months, and linting couldn't catch them.
  • Another asked whether agents can self-promote entries.
• LLMs write too much, one commenter found
  • They built similar tools and found the LLM kept documenting until the system got messy and less useful.
• Run agents short, twice, one commenter argued
  • Agents are probabilistic, so longer runs fail more often.
  • Two 5-minute runs beat one 10-minute run, they said.

How to Apply

• Build gawkbot from source and make a first bot
  • Install Go and Bun, then run the commands below.
  • With a signed-in Claude Code, finish onboarding at http://localhost:7891.
• Add a write-approval gate to your agents
  • Let reads pass, and require user confirmation only for write tool calls like email, commit and delete.
  • List the actions that need approval first.
• Require citations in agent memory
  • Make a source link a required field on every knowledge entry the agent writes.
  • Reject entries without a source.
• Build a per-bot cost view
  • Log dollars, tokens and runs per bot and show them on one screen.
  • Use the cost of one bot build as your baseline.

Try It

# Paste into a terminal (needs Go and Bun / source-build steps from the README)

git clone https://github.com/najmuzzaman-mohammad/gawkbot.git
cd gawkbot
cd web && bun install && bun run build && cd ..
go build -o gawkbot ./cmd/wuphf
./gawkbot --provider claude-code

# The browser opens at http://localhost:7891

Terminology

Related Papers

• Ask HN: If the models are so good, why doesn't OpenAI just start 100 companies?

  HN answers: code is easy, knowing what to build isn't, so they sell tokens instead

• OpenShell – A safe, private runtime for fleets of autonomous AI agents

  OpenShell uses policy files to control agent file, network, and credential access at kernel level.

• Show HN: Omnara – Open-source alternative to Claude Managed Agents

  Open-source managed agent platform runs agents on any model and machine, self-hostable

• Rust off Rails: The opinionated web stack for the AI era

  Sixty Markdown files tell coding agents how to build a Rust web app, with no framework code.

• LayerFS: The filesystem for AI coding agents

  Mount an ephemeral workspace per agent, commit only the changes

• Launch HN: Alter (YC S25) – Access control for multi-user agents

  Alter gives each agent its own identity, policy, and audit trail instead of shared API keys.

• Using Claude to reduce the attack surface of code

  Claude scanned 25 open-source projects and found 493 needlessly reachable code issues.

• WhatsApp's help pages hid LLM instructions: dont recommend a regulator

  WhatsApp help pages hid 102 AI-only instructions, including not recommending a data regulator

Related Resources

• https://github.com/nex-crm/wuphf
• https://github.com/najmuzzaman-mohammad/gawkbot
• https://x.com/karpathy/status/2039805659525644595
• https://gawk.bot
• https://github.com/GistNoesis/Shoggoth.db/
• https://manfre.me/posts/2026/04/build-llm-wiki-obsidian/