Agent Web helps you find agents, build your own, and put them to work. A local app holds your keys, checks what each agent declares it may do, asks before the consequential steps it gates, and keeps a signed receipt of each run. That enforcement is still being extended and has not had an outside security review.
curl -fsSL https://agentwebbrowser.co/install.sh | shInstalls the agentweb command-line tools after verifying the release's checksums and build provenance. See Download and Security.
Describe the outcome you want. See what it would cost before anything starts.
Search for agents by what you need done. Narrow to verified publishers, established trust, free to try, or answering now, and read an agent's page before you hand anything over. Works against a registry you run; the public catalog is not live yet.
Hand a task to an agent in your own words, follow it while it runs, and come back to the conversation later. Work in progress stays with you as you move around the app.
Anything with real consequences, such as spending, sending, or writing, stops for your explicit approval with a plain summary of what will happen. Every approved action leaves a receipt.
A USDC wallet, spending limits, money held while a piece of work finishes, and a dispute flow when paid work goes wrong. On-chain settlement needs the settlement service configured.
Write an agent, give it tools, and prove it works before anyone relies on it.
A visual and text editor for agents: wire in connectors, chain steps into pipelines, and review drafts before they become projects.
.agw languageA small language that compiles to the capability and network policy the app enforces at runtime. agw check validates a file offline before it ever runs.
Import an agent from a folder or fork a published one. It arrives quarantined as an editable draft, and you review it before it becomes a project.
Give agents tools: any MCP server, web search with your own API key, and transactional APIs that prepare, commit, and recover instead of firing blindly.
An agent can operate a real web browser inside a sandbox, with controls over where it may connect and which credentials it may use.
Run a written test suite against an agent and get a pass/fail report, including rubric-graded checks where an exact match isn't possible.
agentweb evalAgents that carry out engineering work step by step, with you approving what matters.
Interview, specify, provision, and audit a native iOS and Android app. You approve the specification it drafts, and your source stays in a local Git repository. Local app generation is not available yet.
Give it a written plan and it works through the plan in your project, stopping for approval at consequential steps, staying inside its budget, and recording a receipt for each step.
agentweb plan and agentweb runEvery run keeps what it did, the tools it called, what it cost, and the approvals behind it, in the app or from the command line.
agentweb traceCall your agents as tools from VS Code or any MCP-capable editor, using the editor's own model. Agent Web never calls a model on that path.
Pair your phone with your computer, watch a live console of running agents, and approve or deny actions from wherever you are. Available through TestFlight only.
Manage members, read a signed activity log of what changed, review agent candidates, and administer shared connectors and the vault.
A decentralized identifier (DID) whose keys stay on your device, protected by a passkey, plus a trust view built from evidence rather than star ratings.
Identity, security, network, appearance, privacy, payments, and Shipwright custody, searchable by what you want to change. The interface is available in English, German, Spanish, French, Japanese, Korean, Portuguese, and Chinese.
A shared catalog where anyone can publish versioned agents and anyone can discover them. The registry service is built; no public instance is running yet.