Read selectors
Capture tables, rows, invoice regions, images, attachments, and chart containers as reliable evidence for the model.
Record the page elements that matter, let your LLM inspect structured evidence, confirm sensitive actions, and create a clean Feishu document report for every run.
Built for reconciliation operators
ReconPilot keeps the human in control while removing repetitive reading, comparing, reporting, and handoff work.
Capture tables, rows, invoice regions, images, attachments, and chart containers as reliable evidence for the model.
Enable Screenshot on any Read selector to send visual context for charts, canvas views, images, and complex page states.
Let the model propose Click/Input actions, then confirm each sensitive operation or switch to auto-approve once the flow is stable.
Create a fresh Feishu document for every run, with readable summaries, evidence, comparisons, decisions, and action outcomes.
Save each reconciliation page as its own task, including page URL, steps, prompts, selectors, scheduling, and execution settings.
Use common providers, local models, or a custom OpenAI-compatible gateway through llm-hub based configuration.
Workflow
ReconPilot runs steps in order. If a step needs confirmation, the next step waits until you approve or reject the proposed action.
Record context. Use Read, Click, and Input to define what the model may inspect or operate.
Analyze with AI. The model receives DOM snapshots, tables, screenshots, attachments, and your prompt.
Confirm actions. Review proposed browser actions before they execute, or enable automatic approval.
Sync the report. Generate a new Feishu document with the outcome, evidence, decisions, and errors.
Use cases
Use ReconPilot anywhere you need structured page review plus auditable reporting.
Compare page-visible amounts, status labels, IDs, images, and exception messages before taking review actions.
Capture visual evidence from charts with Read screenshots, then ask your model to explain trends or anomalies.
Let the extension collect image, text, and document metadata from the page and preserve findings in the final report.
Run configured tasks on Chrome alarms when the browser is open and the target system is already logged in.
Local-first controls
ReconPilot is built as a local Chrome extension. Credentials are stored in Chrome extension storage on your machine, and external calls happen only to services you configure.
FAQ
No. It can run on any normal web page. Each saved task keeps its own page URL, steps, prompts, and selectors.
No. The model returns structured JSON. ReconPilot only executes safe local action types such as click, input, wait, scroll, keyboard, and download.
Each completed run can create a new Feishu document and write a clean Markdown-converted report. Sync can be turned off per settings.
ReconPilot records screenshot and image metadata. If the chosen model or gateway rejects vision input, the result should be marked as needs_review instead of inventing visual facts.
CHROME WEB STORE LIVE
ReconPilot is now officially available on the Chrome Web Store. Click the button below to install the extension and instantly elevate your team's reconciliation efficiency.