</>
Control
Inspect, click, type, extract, and capture network calls in a real browser with structured feedback.
/skill:webcmd-browserSelf-learning browser infrastructure
Stop paying agents to [rediscover] the web.
Webcmd learns websites as agents use them, then compiles that knowledge into faster, cheaper, more reliable commands. Browser-agent token spend drops by up to 90%.
npm install -g @agentrhq/webcmd webcmd skills add # adds the skill to your agent # ask for data, not a browser session
› webcmd hackernews top \
--limit 3 -f json
stories[3]{rank,title,score}:
1, Agent-native interfaces, 418
2, Durable browser workflows, 296
3, Structured tools at scale, 241
✓ completed in 1.2s · 3 results
›
The browser should get smarter every time an agent uses it.
Browser agents are excellent explorers. But repeating a known workflow with screenshots, DOM snapshots, and fresh reasoning is slow, expensive, and fragile.
Webcmd keeps what the agent learns: observed pages, actions, APIs, pitfalls, fallback paths, and finally a durable command with named inputs and structured output.
The loop
Four layers. Less variance at every step.
Start with a real browser. Promote repeated knowledge until the browser is no longer needed.
{ }
Remember
Capture a sitemap of observed pages, states, actions, workflows, APIs, pitfalls, and fallbacks.
/skill:webcmd-browser-sitemap ›_
Author
Build a reusable site adapter with explicit arguments, strategy tags, and consistent output.
/skill:webcmd-adapter-author >>
Execute
Extend the adapter with tailored commands that run instantly with the fewest tokens.
$ webcmd <site> <command> Move known work out of the reasoning loop.
The task Browser agent, every run Webcmd, after learning
Discover the interface Navigate + inspect + infer Known command
Handle inputs Find and fill page controls Named flags
Read results Parse page state Structured output
Repeat tomorrow Start the reasoning loop again Run the same CLI
Illustrative benchmark
Every layer collapses cost and variance.
Modeled repeat-run index for the same known six-step website task, normalized to a fresh browser replay. Directional placeholder, to be replaced with observed benchmark runs.
Input token index lower is better ↓
Fresh browser replay 100
Sitemap-guided 58
Reusable adapter 24
Tailored command 10
Agent turns per run lower is better ↓
Fresh browser replay 9.0
Sitemap-guided 6.0
Reusable adapter 2.0
Tailored command 1.0
90% lower token index 8× fewer agent turns 14× faster task execution
One small surface. Many execution strategies.
Webcmd chooses the cheapest reliable path: public endpoints, authenticated cookies, intercepted requests, page UI, or a local tool. Agents discover them all the same way.
See how commands are built agent discovery webcmd list -f json
commands[5]{site,command,strategy,browser}:
hackernews,top,PUBLIC,false
twitter,whoami,COOKIE,true
linkedin,safe-send,INTERCEPT,true
reddit,comment,UI,true
docker,ps,LOCAL,false
help[2]:
webcmd <site> --help
webcmd <site> <command> -f jsonSpend agent attention on the task, not the navigation.
Structured by default
Every interaction returns machine-readable matches, confidence, output, and errors.
Strategy-aware
Use PUBLIC, COOKIE, INTERCEPT, UI, or LOCAL depending on the cheapest reliable path.
Agent-discoverable
One list command exposes arguments, output columns, browser requirements, and strategy.
Start here
Make the browser learn.
Install Webcmd, inspect the command registry, and fall back to live browsing only when you need it.
npm install -g @agentrhq/webcmd webcmd skills add # adds the skill to your agent