LibreAI / KI-Agenten / webcmd
// agents · agentrhq/webcmd

webcmd

webcmd ist ein Open-Source-Projekt aus der Kategorie KI-Agenten von agentrhq (agentrhq/webcmd auf GitHub). Es hat 2.665 Sterne auf GitHub und 1.542 Forks, wird unter der Lizenz Apache-2.0 veröffentlicht und wurde zuletzt am 25. September 2026 auf GitHub aktualisiert.

Self-learning browser infrastructure for AI agents — sites remembered, so agents stop rediscovering the web

★ 2,7 Tsd. Sterne⑂ 1,5 Tsd. ForksApache-2.0aktualisiert Sep 2026
▦ Visual — das README dieses Projekts enthält keinen einbettbaren Screenshot,
darum hier ein Platzhalter. Die Anleitung unten ist trotzdem vollständig.
// projektdaten

Projektdaten

Repository
agentrhq/webcmd
GitHub-Sterne
2.665
Forks
1.542
Lizenz
Apache-2.0
Zuletzt auf GitHub aktualisiert
25. September 2026
Kategorie
KI-Agenten
Eintragstyp
Vollständige Installationsanleitung
Offizielle Website
Official site
// so funktioniert es

So funktioniert es

  1. Clarify your goal

    Name the outcome you want from the web — research with source links, data from a logged-in site, a price/availability check — in one prompt.

  2. Get Webcmd

    Install the npm package globally (Node.js 20.6+ required).

  3. Install the skill

    Run webcmd skills add and choose your harness (Claude, Codex, another supported harness, or a custom path).

  4. Run your first task

    Tag the webcmd-browser skill for live browser work and describe the outcome; Webcmd works in a real browser and quietly retains navigation context for next time.

// funktionen

Warum es so beliebt ist

Live browser control

Agents inspect, click, type, extract and capture network calls in a real browser to complete tasks on unfamiliar sites.

Sitemap memory

For familiar sites, Webcmd keeps an agent-facing sitemap of observed pages, states, actions, workflows, APIs, pitfalls and fallback paths.

Logged-in profiles

Profiles are cookie jars for authenticated sessions (e.g. a 'work' or 'social' profile); sessions are independent browser windows within a profile, so parallel agents can work safely.

Sandboxed browser programs

For multi-step local exploration, agents can send one Playwright-style program to an explicit browser session via webcmd browser run.

Learning stays local and quiet

After an optional first-access cloud seed, subsequent learning stays local; learning never blocks the task and the live browser is always the truth.

Broad site coverage

Used across research/community sites, social and professional networks, AI tools, shopping and booking products — the README stresses the same workflow works on other sites too.

Advertisement
// selbst ausführen

Führe es selbst aus

  1. Clone the repo

    Clone webcmd if you want the source; the normal install is the npm package in step 2.

    git clone https://github.com/agentrhq/webcmd.git
    cd webcmd
  2. Install Webcmd

    Install globally and add the browser skill to your harness, as the quick start shows.

    npm install -g @agentrhq/webcmd
    webcmd skills add
  3. Run a browser session

    Create a named session in a profile, inspect its tabs, and close it when done (raw browser commands need an explicit session ID).

    webcmd --profile work session create "Work Project" -f json
    webcmd --profile work session close work-project-k7
// anwendungsfälle

Was du damit bauen kannst

Advertisement
// häufige fragen

Häufige Fragen

Who maintains Webcmd?

agentrhq (AgentR) maintains it; docs are at webcmd.dev.

Is Webcmd free?

The package is Apache-2.0 licensed open source. Webcmd Cloud (hosted sessions) is described as in active development and not yet stable.

Does it need a paid API key?

Webcmd is browser infrastructure your agent drives; model costs belong to the harness (Claude, Codex, etc.) you connect it to.

How does it compare to a plain Playwright script or browser MCP?

Those start from zero on every run. Webcmd adds a persistent, self-learning memory of how sites work, so repeat tasks skip rediscovery.

How do I get started?

npm install -g @agentrhq/webcmd, run webcmd skills add, then ask your agent for a web outcome using the webcmd-browser skill.

// weiter entdecken

Ähnliche Tools

KI-Agenten

steel-browser

A batteries-included browser API so agents can automate the web.

★ 7,8 Tsd. SterneApache-2.0
Anleitung ansehen →
KI-Agenten

PraisonAI

PraisonAI 🦞 — Hire a 24/7 AI Workforce. Stop writing boilerplate and start shipping autonomous self-improving agents that research, plan, code, and execute t…

★ 9,1 Tsd. SterneMIT
Anleitung ansehen →
KI-Agenten

reef

Reef — infrastructure for continually self-improving agents: weight-training and harness-evolving deployments.

★ 7,5 Tsd. SterneApache-2.0
Anleitung ansehen →
KI-Agenten

pentagi

Autonomous AI agents for penetration testing — self-hosted, multi-agent, with a full web UI

★ 25,3 Tsd. SterneMIT
Anleitung ansehen →
KI-Agenten

brightdata-mcp

Bright Data's MCP server: web search, scraping, structured data and browser automation for agents (75 tools)

★ 2,7 Tsd. SterneMIT
Anleitung ansehen →
KI-Agenten

stop-that-shit

Hook and skill guard that stops coding agents from doing unrequested busywork.

★ 2,5 Tsd. SterneMIT
Anleitung ansehen →
// alternativen

Alternativen im Bereich KI-Agenten

Die anderen Projekte der Kategorie KI-Agenten mit den meisten Sternen im LibreAI-Katalog — die realistische Auswahl, falls webcmd nicht das Richtige ist.

// mehr in agents

Mehr in KI-Agenten

Verwandte Projekte der Kategorie KI-Agenten im Katalog, die webcmd bei den Sternen am nächsten stehen — zum Weiterstöbern auf beiden Seiten.

Hol dir webcmd

Kostenlos und offen — hol dir den Quellcode, gib dem Repository einen Stern und leg los.