A browser agent that fills web forms from your documents — and asks you when the document is not enough.
-4285F4?logo=googlechrome&logoColor=white){kind=icon}
- Download
koni-forms-1.0.0.zipfrom Releases and unzip it. - Open
chrome://extensions, turn on Developer mode, click Load unpacked, and select the unzipped folder. - Click the KONI icon, open ⚙ Settings → Models, add your provider and API key, and pick a model for the Planner and the Navigator.
- Open any web form, attach a DOCX or PDF with 📎, and tell it to fill the form.
A step-by-step guide with screenshots, model requirements, and troubleshooting is in the release notes.
Filling a web form from a document is mostly transfer work: find the value, type it into the right field, repeat. KONI-Forms adds a document layer to a browser agent so it can work from the document instead of guessing, and a user-intervention loop so that a value the document does not settle becomes a question, not a fabrication.
- Document-grounded filling — search the document, read its tables, view PDF pages, and point a form field at the exact document cell it came from.
- Asks instead of guessing — when the document does not settle a value, the agent asks you and continues with your answer; when it cannot continue, it stops and says why.
- You stay in control — confirmation before irreversible actions (submitting forms, payments, sending messages), and optional plan approval before a run starts.
- Your own API key — OpenAI, Anthropic, Gemini, Groq, Cerebras, DeepSeek, Grok, OpenRouter, Azure OpenAI, Llama API, and OpenAI-compatible endpoints.
The evaluation fills three public OSWorld 2.0 web forms from documents and scores the saved form state against gold answers.
| Form | Site | Tasks | Field modes in gold |
|---|---|---|---|
| Expense claim | ExpenseFlow | expense_01 – expense_03 |
AUTO, ASK_USER |
| Insurance claim | InsClaim | insurance_01 – insurance_03 |
AUTO, ASK_USER, USER_ONLY, MUST_BE_EMPTY |
| Account registration | CloudCRM | crm_01 – crm_03 |
AUTO, ASK_USER |
Each of the nine tasks comes at three document scales, and each scale as both DOCX and PDF — 27 variants in all:
| Suffix | Pages | Document |
|---|---|---|
| (none) | 1–2 | the original document |
x1 |
28–36 | the same document surrounded by policy text, filing guidance, and a specimen form carrying the same labels |
x2 |
44–57 | the same, with more of that surrounding material |
Target values and gold states are identical across the three scales: only document length and distractor content change, not the task. Page and word counts for every variant are in experiments/docs/_manifests/.
Gold annotates 191 AUTO fields (CRM 36, insurance 43, expense 112). Fields that must stay empty, expected user questions, and user-only steps are annotated separately and lie outside the 191. The scorer is a single standard-library Python file:
import json, score
gold = json.load(open("gold/insurance_01.json", encoding="utf-8"))
score.CFG = score.DOMAINS[gold["domain"]]
print(score.score(state, gold, asked=[], on_page={})) # state: the site's /api/state JSONSite setup, the gold format, and the scoring rules are in experiments/README.md.
Requires Node.js 22.12+ and pnpm 9.15.1.
git clone https://github.com/Byun11/KONI-Forms.git
cd KONI-Forms
pnpm install
pnpm build # unpacked extension in dist/
pnpm zip # zipped extension in dist-zip/chrome-extension/– background service worker: planner/navigator agents, document tools, browser control.pages/– side panel (chat, document attach and parsing), options page, content script.packages/– shared storage, i18n, UI components, and build tooling.experiments/– documents, gold answers, OSWorld 2.0 site configs, and the scorer.
Apache License 2.0 — see LICENSE. Based on Nanobrowser (Apache-2.0).
Security issues: please use a GitHub Security Advisory.
Developed at KISTI (Korea Institute of Science and Technology Information).