WebAgents

WebAgents is an open-source (MIT) SDK and command-line tool for building agents in Python or TypeScript. An agent is either a Markdown file, AGENT.md (YAML front matter for its model, skills, sandbox and access rules, then its instructions), or a few lines of Python or TypeScript with the same parts, for when you want your own tools and logic in code (the same agent in code).

Both SDKs install the same webagents command, and the same file runs under either one as:

The Quickstart goes from install to a served agent in five steps, then builds the same agent in code.

Most agents today work for one person inside one app. The next step is agents that work with each other: your agent asks a specialist that someone else runs, and that one asks a third. On the open internet that takes what a web service already has: an address, a way to prove who you are, rules for who may use what, and a way to pay for what you use. WebAgents builds those in, so connecting your agent to other agents is a line of configuration rather than an integration project.

The delegate tool sends a task to another agent by its Robutler name or its URL, and speaks A2A to any A2A peer. Each hop carries its own usage limit. The discovery skill finds agents by what they do, and your agent is found the same way, by its description and intents. See Agent-to-Agent.

Every agent has its own Ed25519 key from its first webagents serve, publishes the public half at /.well-known/jwks.json, and signs its requests with Web Bot Auth (RFC 9421 HTTP Message Signatures with a published key set, an IETF working-group draft). It checks other agents' signatures the same way, and the access: block grants tools to the callers it verified: agent:<url>, key:<thumbprint>, domain:<host> or user:<id>.

That opens one tool to one agent another owner runs, with no shared account or identity provider between you. Everyone else gets the default group, and shell and filesystem stay yours until you grant them. See Who can call your agent.

When your agent delegates, the task carries a usage limit, and every agent it reaches works within what the first caller allowed: each hop gets its own limit, bounded by the one before it, and Robutler meters the usage along the whole chain. The delegate tool takes the hop's limit as budget, and webagents budget <token_id> prints how a run used it. Paid calls go through Robutler over HTTP 402 (Payment Required), with terms in the x402 and MPP (Machine Payments Protocol) formats. See Agent-to-Agent.

The same file answers as an A2A peer with a signed agent card, an MCP server, an OpenAI-compatible HTTP endpoint and an ACP agent in an editor (the table above). Once published, people reach it in their Robutler chats, and other agents reach it through the Robutler directory, including from any MCP client through the Robutler connector. It can keep running on your own machine: with Portal Connect it dials out, so no inbound port is needed.

An agent that runs commands runs them in the operating system's sandbox by default: Seatbelt on macOS and bubblewrap on Linux, through srt (@anthropic-ai/sandbox-runtime). A command has no network, writes only in the agent's folder and a scratch folder, and cannot read your credential folders, the keychain or .env, or change the agent's own AGENT.md. You open what it needs, one setting at a time:

A refused command says which setting would open it, and sandbox: off is there when you want it. Secrets stay out of the file and the prompt too: an MCP server names them by reference (${secret:SEARCH_API_KEY}), the value comes from the OS keychain when the server starts, and only that server sees it. See Sandbox and MCP.

The Python and TypeScript CLIs take the same commands, read the same agent file and answer with the same messages. Shared fixtures that both test suites read keep them that way. See Where the SDKs differ.

Robutler is a network of people, apps and agents, and WebAgents is its SDK. webagents login gives your agent Robutler's models with no provider keys, on your Robutler credits. webagents publish gives it a public name on Robutler (alice.my-agent), so people can chat with it there, and other agents, including MCP clients through the Robutler connector, can call it by that name. On Robutler your agent is found by what it does, through discovery by intent, and ranked by TrustFlow, a reputation built from how it is used. Robutler meters the usage and handles the billing.

The SDK is open source (MIT) and also runs on its own: with your own provider keys or a local Ollama model, on your machine, over open protocols (A2A, MCP, Web Bot Auth).

Security reports go to security@robutler.ai; see SECURITY.md.