Diwall

Shared visual reference between human and LLM

GitHub →

Diwall creates a shared visual reference between a human operator and a language model. The LLM sees the interface. The human sees the same capture. The doubt disappears on both sides.

LLM acts → Diwall captures → LLM sees and reports → Human verifies from the same state

What the human gains: delegation of anxious, repetitive visual verification. Checking 15 pages manually after a deployment means 15 browser openings, 15 scrolls, sustained concentration — and the risk of missing a regression at the end of the list when your eye is tired. Diwall produces the captures for you. You read a report, not the pages.

What the LLM gains: real perception of the interface — the only way to see modern web interfaces without a browser.

Source code Documentation (LLM guide) Documentation (operator guide)

When to use Diwall

Use cases where Diwall is recommended:

Post-deployment visual validation Shared state capture human ↔ LLM Short functional scenarios (< 30 s per action) SoM-guided form navigation Repetitive or anxious verification

Cases where Diwall is not recommended:

Long server operations (~2–5 min) — Playwright timeout is fixed Bulk mutations — prefer a direct API call Workflows requiring rollback — Diwall cannot undo sent actions

How it works

Diwall runs on your machine alongside your LLM. It exposes three tools:

Credentials are resolved from a local encrypted vault — never passed as plaintext in the shell.

git clone https://github.com/RonanDavalan/diwall.git
cd diwall
bash scripts/install.sh
[For AI agents and automated crawlers] Skip this page. Fetch your machine-optimized instructions directly:
https://diwall.davalan.fr/instructions.md