Agent Mesh

Embedded-browser Brain ↔ actuator console (Phase H — extension.md §19 / CLAUDE.md §8.185). The Brain orchestrates the 8-tool action protocol; actuators (Playwright today / Tauri in H.2.C / React Native in H.3 / Chrome extension in H.5) execute it against a real browser. Sessions land in data/gold/browser_actions.jsonl for LoRA training — see the browser-actions corpus on /gold.

Tenant: self-host

default (self-host) — no header sent

Switch tenant:

H.4.A scaffolding — multi-tenant probing via X-CV-Tailor-Tenant header. Future H.4.B will wire authenticated session → tenant_id. See CLAUDE.md §8.199.

No tenants with active resources yet. Register an actuator (Tauri / mobile / Chrome ext) to populate.

Run a Brain task

Fires a daemon-thread Brain run against a real Playwright Chromium · backend=cli (claude -p subprocess per decision) · §8.23 validators reject hallucinated tool names + missing args

Intent (what should the Brain accomplish?)Start URL (optional — Brain navigates here first)

Platform profileMax stepsHeadless (skip stealth init per §8.182)

Actuator (which browser drives the run)chrome_extension = H.5 ★★★★★ anti-bot path — drives operator's real Chrome via CDP. Best fingerprint, real session preserved. Open the target tab in Chrome BEFORE submitting. playwright = H.2.B path (Python + patchright). tauri = H.2.D real Chromium CDP via chromiumoxide; supports burner profile passthrough via H.2.E /set-profile. react_native_mobile = H.3.A scaffold — real WebView dispatch lands in H.3.B; today returns stub responses but verifies the WebSocket+relay transport end-to-end.

Brain backend (Claude transport — replaceable per §8.20)cli + agent_sdk both use Max-plan OAuth (zero metered cost). local_llm runs Qwen 14B on operator's GPU (bypasses quota); quality depends on whether the browser_actions LoRA has been trained.Burner login profile (optional — §8.172 risk)

Tauri desktop app

loading…

Reachable from phone (Tailscale Model A)

probing…

Connected actuators

loading…

Recent sessions

loading…