htmlllm is a human-first kind-of single-file kind-of harness with a focus on writing and understanding.
If you're tired of reviewing what agents write, turn the tables: let agents review what you write.
Writing clearly, and most importantly, understanding, takes time and no amount of innovation in LLMs will change that.
To be a bit tongue-in-cheek: don't get caught again saying "I don't know, the agent wrote it". Write what you know, ask questions and let agents help with a more thorough understanding. Repeat.
- simple: no server, database, API or setup, just a simple markdown-based block editor communicating with a session on your machine
- agent is a reviewer: posts comments on paragraphs, replies to your comments
- brief vs. explanatory mode: the style of agent's comments
- a basic work-in-progress stack for monitoring ongoing agentic work
- basic export: click Export, it opens a panel with complete markdown (content only, not revision comments); copy it and paste it where you need it
- chromium-based browser (Chrome, Brave) or Edge with File System Access API enabled
- Claude with Monitor tool (built-in)
HTML file is the shell, json file is the yolk.
When you perform an action (add a paragraph, edit paragraph, create a comment, reply to a comment, resolve a comment), it is written to the json file. Agent has a monitor on that file and watches for changes. Once it detects the change, the agent parses the change and decides what to post as a comment (or not to post a thing).
- clone this repository; navigate to the directory
- make a copy of
copyme.html, give it a descriptive name - open a file in Chrome/Edge
- click on Start a new doc and create an associated json file; give it a descriptive name
- start a new interactive agent session in the directory where
AGENTS.mdis (current directory if you didn't change it) - send it a message ("read AGENTS.md") and agent will ask you which files to watch; choose the newly created pair
- start what you do best
You can also run this in a separate project if you want to use that specific project's memory. In that case, copy the copyme.html, theme.css (if you want theming) and point an agent to the htmlllm's AGENTS.md so it knows what to do. If you plan on using it continuously in another repository, it's advisable to integrate htmlllm's AGENTS.md / CLAUDE.md into that project's AGENTS.md / CLAUDE.md.
This was tested mainly with Claude. With an agent that lacks a monitor-like tool you may have to get creative, or just use something else.
Can it use an API instead of local session? Yes, probably, but my focus was on almost zero-setup.
Although this can be used for regular agentic work (like coding, research etc.), the granularity of what's going on is limited, compared to what you would actually see in a session. You get a basic view of work in progress and that's it.
MIT — see LICENSE.