agentweb-cli 0.33.1

Turn websites into fast, structured commands for coding agents

A browserless web interface for coding agents — typed commands in, JSON out.

Docs  ·  Supported sites  ·  Build an adapter  ·  Security

Both commands are real, keyless and read-only. The recording is scripted, not narrated: docs/assets/demo.tape regenerates it with vhs docs/assets/demo.tape.

---

What it is

A website is compiled once into a typed, versioned adapter. After that, agents replay its mapped operations as ordinary HTTP requests that return bounded structured JSON. The browser is an authoring and login surface, not the ordinary execution engine.

That moves finding controls, filling forms and interpreting results from every agent run to one shared adapter. It also makes the limitations explicit: AgentWeb only works on sites and operations someone has mapped, and a typed operation is not automatically proof that the whole user task works.

Install

macOS and Linux

If Python 3.11 or newer is not already available, the installer provisions an isolated Python for you.

curl -fsSL https://github.com/AnayGarodia/agentweb/releases/latest/download/install.sh | sh

Restart Claude Code or Codex, then paste this prompt:

Use AgentWeb to find the latest version of React on npm.

Everything lands under ~/.local/share/agentweb, plus a global AgentWeb skill for Claude Code and Codex holding the CLI's absolute path — so new agent sessions find it even when the app does not inherit ~/.local/bin. It does not change your system Python packages and does not register MCP automatically.

Prefer to read scripts before running them? Read install.sh, or install from a checkout with git clone, python3 -m pip install -e ., and agentweb setup. Already installed? agentweb upgrade re-runs the official installer after checksum-verifying it, preserving your sessions and profiles; agentweb upgrade --check only reports installed vs. latest.

Windows

Run the signed release-channel PowerShell installer:

irm https://github.com/AnayGarodia/agentweb/releases/latest/download/install.ps1 | iex

A one-line PowerShell installer, the counterpart of install.sh, is built from installer/install.ps1.template and published with the same immutable release. Read it before you trust it: it has not been run on a real Windows machine. Its argument handling, wheel verification, Python discovery and every failure message are covered by tests that execute it under pwsh on Linux and macOS, which catches the class of bug that broke it once already — but that is not a verified Windows install, and we will not claim it is until someone runs it on Windows. Reports welcome.

There is no PyPI release yet, so pip install agentweb-cli does not work on any platform. The release installer above is the current answer.

Just ask for what you want

Prompts are ordinary tasks. The agent handles discovery and command syntax.

Use AgentWeb to find three recent arXiv papers about coding agents.
Use AgentWeb to compare the dependencies of React and Vue on npm.
Use AgentWeb to research Alan Turing on Wikipedia and follow the most relevant links.

If the installer could not detect your coding agent, connect it by hand with agentweb install-agent claude --scope user or agentweb install-agent codex --scope user. See agent setup for the exact behaviour hosts should follow around login, retries, and writes.

What the agent runs underneath

agentweb sites                                            # what is installed
agentweb capabilities arxiv.org                           # what one site supports
agentweb describe arxiv.org --operation search_papers     # one operation's contract
agentweb arxiv.org search-papers --query "attention" --limit 3

Reads look things up. Acts do things on your own accounts — play a song, add to a cart, open a pull request — and impactful or irreversible ones require an explicit --confirm. Actions are the point: reads are often a web search away, but doing things is what a browserless interface uniquely unlocks. Commands accept plain domains and most website URLs, and a command's main argument works positionally or as a flag, so copy-pasted examples work either way. Results and expected errors are JSON, so agents need no site-specific parsing code.

Why not Playwright MCP?

Microsoft's own Playwright MCP README makes the general case for the CLI shape:

Modern coding agents increasingly favor CLI–based workflows exposed as SKILLs over MCP because CLI invocations are more token-efficient: they avoid loading large tool schemas and verbose accessibility trees into the model context, allowing agents to act through concise, purpose-built commands.

— microsoft/playwright-mcp

AgentWeb pushes that one step further: the mapping is done ahead of time and versioned, so the agent never reasons about page structure at all. Browser automation is still the right tool for a site nobody has mapped and for exploratory work — and AgentWeb ships agentweb mcp if you would rather drive it through MCP anyway.

The MCP surface is five tools and stays constant as adapters are added: site_find, sites_list, site_describe, site_call, and site_connect. The CLI remains the lower-context path.

What has actually been measured

In the current live task-completion run, AgentWeb completed 10/10 multi-step read tasks it attempted; a stepwise local Chrome completed 4/10. That is a completion result, not a speed or cost win. Both arms followed fixed shortest paths with no model in the loop. This page makes no speed or cost claim.

Writes were 0/0. No authenticated write ran in this benchmark, so it is not evidence that carts, playlists, votes, or issues work end to end. Two separate signed-out refusal checks are reported separately and are not counted as completed actions. Method, task table, and limitations.

Supported sites

30 sites, 364 mapped operations in the signed offline CLI bundle, with signed adapter updates available from the hosted registry. On the product denominator — reviewed meaningful user tasks — 0 of 30 currently earn the Verified tier. Operation evidence and route coverage are reported separately. See all sites and operations, or run agentweb sites for what is installed locally.

Before signed sync, the bundled non-mapping runtime exposes 483 operations after shared fallbacks. Live agentweb sites is dynamic and can report more after a signed hosted adapter upgrade.

Honest coverage

Adapters state only what they have actually proven. Every operation carries a status computed from bundled evidence rather than asserted by a maintainer: verified (a live replay whose content was checked against reviewed assertions), shape-verified (a live replay that matched the reviewed schema), or unverified (mapped and typed, but with no machine-checked evidence bundled). Adapters also declare their known gaps, so an agent can check before promising that a site is fully covered.

Coverage stays honest over time, too. Any operation can be captured once as a response oracle — the redacted response shape plus optional assertions, never captured values — and replayed later to prove the site has not changed underneath the adapter. agentweb verify-oracles --dir oracles --strict replays a whole directory and exits non-zero on drift, which is what the scheduled Oracle drift workflow runs against the committed oracles/. The full mechanism, including the browser-assisted read-back path, is documented in architecture and the adapter guide.

Login only when needed

Try public commands without logging in. When an account action needs authorization, AgentWeb returns authentication_required and you run agentweb connect example.com. That window opens already signed in by default, copying only the target site's session files from your everyday Chrome profile — never your history or extensions, and never cookies shared across sites. Pass --isolated for a blank window instead. Sessions stay on your device under ~/.agentweb and are never bundled into an adapter. Read security and trust before installing an authenticated third-party adapter.

Add a site

Mapping a public, read-only site is roughly a thirty-minute job:

agentweb adapter-new example --base-url https://example.com --root ./registry
agentweb registry-build ./registry
agentweb audit example --root ./registry

Scaffold the adapter, describe the endpoints you want as typed operations, build the index, then audit — audit verifies adapter contracts and declared full-site coverage, so it will tell you when you have claimed more than you mapped. Building an adapter covers login, write confirmation, evidence, and what must never be committed. Would rather someone else map it? Request a site.

What is in this repository

This repository is the Apache-2.0 open core: the runtime, CLI, adapter tooling, and the public adapters listed above, bundled as an offline capability floor. agentweb sync can add signed hosted adapters without removing that release bundle. The automated mapping workbench is not public. Official adapter bundles and hosted services may use separate terms.

Usage analytics

agentweb dashboard opens a private dashboard on 127.0.0.1 showing installs, activation, repeat use, success rate, latency, and failures for this installation. AgentWeb never records prompts, arguments, URLs, website responses, account identities, cookies, or credentials — run agentweb telemetry inspect to see exactly what one event contains, or agentweb telemetry disable to switch it off (details).

Status

AgentWeb is an early public preview. Websites change endpoints, revoke sessions, add verification, and rate-limit requests; AgentWeb reports an adapter's known gaps and verification state instead of claiming every action will work forever.

Contributions are welcome — see contributing, architecture, and the changelog. Coding agents working on this repository should read AGENTS.md; llms.txt is the machine-readable documentation map.