Over-ambitious
Client game + websocket + server-side resolver + FSM + chips + accounts + WebMCP extension. And the house bot. Most of it was up, and it was still too much.
A pivot story · 2026
From AgentHoldem
to Auto Browser.
A pivot story, and what I built from its ashes.
Vin Lim · autobrowser.dev
01 · The idea
I set out to converge
I set out to converge
AI Agents · WebMCP · Bitcoin.
After running through many ideas I settled on AgentHoldem, a hackathon project where agents play Texas Hold'em against each other, driven by WebMCP, paid in Bitcoin.
Agent Warfare
Agents playing each other. Sounds like fun.
A WebMCP Testament
Fast-paced play demands structured tools, not DOM scraping.
Bitcoin-Native House
Poker + crypto is a natural fit. Blocknomics-friendly by design.
02 · The artifact
This was AgentHoldem.
Live at agentholdem.xyz. Agents battle at the table, a reasoning log on the right shows every thought and action. WebMCP under the hood.
agentholdem.xyz
03 · The pitch I told myself
It felt like a cheat code.
Gaming is addictive. If I could get people to play, traction would take care of itself.
Three trends converging. Demo-ready hackathon scope. A narrative that writes itself. What could go wrong?
04 · Second thoughts
Three realities broke it.
Security posture
Real money in a hot wallet attracts the wrong attention. I wasn't ready to defend it, and shipping something I can't defend is worse than not shipping.
Traction paradox
Gaming is addictive until someone else plays for you. Agentic poker removes the fun, especially when you're paying tokens for the agent to have it.
05 · The hard pause
“I'm almost all in. If I pivot, I lose time for traction. If I carry on, I'm unlikely to get traction anyway.”
Sunk cost is real. So is the cost of the wrong road.
06 · The pivot
So I doubled down on
So I doubled down on
my favorite part.
Of the three pillars, one was a feat of engineering on its own: the WebMCP Agent Orchestrator for Chrome. I couldn't stop building it. So I stopped building everything else.
AgentHoldem Auto Browser
Lesson one
Hackathons reward the
Hackathons reward the
smallest thing with a real story.
Not the most ambitious thing half-built. The AgentHoldem stack had ten working parts but no single one I could demo without apologising for the other nine.
Lesson two
Don't ship what you
Don't ship what you
can't defend.
Security posture isn't optional, even for demos. The moment real money is in reach, you're the adversary's target. If you're not ready to own that, the right answer is to not handle money.
Lesson three
Pivot toward the part you
Pivot toward the part you
can't stop building.
Not the most valuable part on paper. Not the part closest to traction. The part your hands keep returning to when nobody's watching. Of AgentHoldem's three pillars, only one survived contact with the second-thought.
Part two
Auto Browser.
An AI agent that actually uses your browser.
07 · What it is
A Chrome side-panel
A Chrome side-panel
agent, WebMCP-first.
- WebMCP-first: page-authored tools beat DOM scraping, every time.
- Privacy-first: no telemetry, no analytics, no backend server.
- On-device options: local LLMs or Chrome's built-in AI.
- Four providers: one unified interface across cloud and local.
Auto Browser Sonnet 4.6 Press Run to watch the agent think.
08 · Architecture
Five moving parts.
Sidebar UI user surface
Content script ReAct loop
Provider layer 4 LLMs, unified
Service worker per-tab state
WebMCP shim document_start
No backend. The service worker holds per-tab state in
chrome.storage.local; everything else runs in your browser.
09 · Orchestration
Five-phase ReAct.
plan
Break the goal down into 1–3 concrete next steps.
analyze
Read the current page state and pick what applies now.
act
Call one tool (WebMCP first, then CDP) and run it.
verify
Did it actually take effect? Check with read-only tools.
recover
Something went wrong. Diagnose, replan, or bail cleanly.
Each turn emits one decision object. The phase field tells the orchestrator what to do next. No free-text loop, no regex parsing.
// src/llm/react-schema.js { "phase": "plan" | "analyze" | "act" | "verify" | "recover", "narration": "≤15-word user-facing label", "tool": "tool_name" | null, "args": { /* tool arguments */ }, "final_answer": "task_complete prose" | null }
10 · Memory
Working memory isn't
Working memory isn't
compression.
Think of it like a hiker's logbook. Before every step the agent checks: where am I, where have I been, what worked, what failed. So it doesn't walk into the same wall twice.
Most agents just shrink chat history to fit. We keep a live ledger of observations, routes, and attempted actions.
Where am I? Each page/state is fingerprinted so the agent recognises it if it comes back.
Where can I go? Paths that are open, blocked, or already visited.
What have I tried? Every action, its result, and how many times it's been repeated.
Am I going in circles? Bail after 3 identical failures and try something else.
Result: fewer wasted turns, smaller prompts, better decisions. src/llm/working-memory.js
11 · The USP
Four providers, one interface.
OpenRouter
Cloud · 200K ctx
Access to reasoning models across Anthropic, OpenAI, Google. Default pick.
Gemini
Cloud · 1M ctx
Extended thinking via thinkingConfig. First-class for long tasks.
Local OpenAI-compat
Your server · Ollama, vLLM
Bring your own model, your own weights, your own network. Zero egress.
Chrome Built-in AI
On-device · Gemini Nano
Private by construction. Multimodal Blob pass-through, no base64 hop.
One LLMProvider contract:
availability · call · supportsModality · dispose.
Tool calls, streamed tokens, and reasoning (native or
schema-faked) all surface through the same onThought
callback. The ReAct loop doesn't know which one answered.
12 · WebMCP-first
Why we're WebMCP-first.
WebMCP gives the agent a typed contract instead of a screenshot to guess from. That single choice shapes everything Auto Browser does.
Contracts, not screenshots
Page authors register tools with real parameters, like
addToCart({sku, qty}). The agent calls them
directly. No guessing which button is "Add to cart".
Fast enough for games
Parsing the DOM or a screenshot is slow. Calling a typed function isn't. Tool turns cost 20–100 tokens instead of 2000+, so the agent replies in a fraction of the time. Fast enough for agent-vs-agent games, live dashboards, anything where latency matters.
Survives UI changes
Your CSS refactor doesn't break the agent. Your A/B test doesn't either. The tool contract is the stable layer, whatever happens to the pixels.
Auditable by design
Structured args in, structured results out.
requestUserInteraction() pauses for consent on anything
irreversible. Every call is a log line you can read.
13 · Toolkit
30 tools ship in the box.
When a page has WebMCP, its tools come first. When it doesn't, Auto Browser falls back on a CDP-backed toolkit, so the agent is never stuck. Every tool is permission-gated: read-only calls run silently, mutating ones ask once per domain.
Perception 5
- take_snapshot
- take_screenshot
- get_page_text
- find
- computer
Actions 7
- click
- hover
- fill
- fill_form
- press_key
- scroll
- drag
Nav & Debug 11
- navigate
- go_back
- go_forward
- reload
- wait_for
- wait_for_network_idle
- list_network_requests
- get_network_request
- read_console_messages
- evaluate_script
- get_element_info
Emulation 4
- set_viewport
- set_user_agent
- clear_emulation
- handle_dialog
Orchestration 3
- ask_user
- ask_user_form
- set_poll_interval
WebMCP tools from pages shadow built-ins with the same name. Page-authored contracts always win.