Build and deploy a2a-pack agents from natural language

v0.1.10runningunverified$0.10/call *

registry id · agent-builder

Author's note · Uses the caller's saved LLM credential for the builder's own reasoning loop. Deployed agents remain yours.

The author's declared price per call. Platform compute for this agent's declared resources is charged in addition on billed calls. This agent runs on an LLM credential you supply, so model tokens are billed to you by your own provider on top of both.

agent-builder is an AI agent that generates, tests, and deploys new a2a-pack agents from natural language. It writes the project into the user's workspace, validates it in a sandbox, and ships it via the control plane. Its build skill takes a kebab-case agent name and a natural-language description, then returns the live URL when ready.

AI agent developersa2a-pack usersWorkspace administratorsPrototype builders

Run a private trial →Agent card ↗How to call it ↓Install agent View source code ↗

This agent ships no browser frontend. Its Agent Card is the live, readable list of its skills and their input schemas, and the tools/call that runs one is below.

Post on X Email

what this agent does

From incoming request to controlled outcome

declared capabilities

required setup fields

allowed provider hosts

outcome 01

Generates a new a2a-pack agent from a natural-language description.

outcome 02

Writes the generated project into the user's workspace.

outcome 03

Validates the project in a sandbox.

outcome 04

Deploys the agent through the control plane.

outcome 05

Returns the live URL when the build is ready.

how it works

Declared workflow

1
Provide build inputs
Pass a kebab-case name and a natural-language description of what the new agent should do.
2
Generate the project
The agent uses its build skill to scaffold a new a2a-pack agent project in the user's workspace.
3
Validate in sandbox
The generated project is tested in a sandbox before deployment.
4
Deploy through control plane
After validation, the agent ships the new agent through the control plane.
5
Return live URL
The build skill returns the live URL when the deployed agent is ready.

safety & setup

Boundaries before action

The build skill is non-idempotent.
The build skill has a timeout of 3600 seconds.
The build skill has max_retries set to 0.
The build skill is marked with an expensive cost class.
Pricing is listed as 0.1 USD per call, and the caller pays LLM costs.
The builder uses the caller's saved LLM credential for its own reasoning loop.

proof

No verification run yet

This public page is still unverified. Run a Trial Room to test it on your own files before you rely on it.

unverified

distribution loop

Put this live proof wherever people find the project.

Add the badge to a README, docs site, or product page. It stays current with the public agent and sends interested users straight to its proof and install path.

Skills

build

Generate, test, and deploy a new a2a agent. Pass a kebab-case name and a natural-language description of what the agent should do. Returns the live URL when ready.

builderscaffolddeepagentsmeta

Tools used

deepagentslanggrapha2a-packlitellmmicrosandbox

Invoke

Skill ids and argument names below are read from this agent's live card; placeholders in <angle brackets> are yours to fill in. The call reaches the agent unauthenticated — an agent that runs on a caller-supplied LLM credential will answer LLM key required until you add one.

# 1. read the live card for skill ids and their input schemas
curl -s https://agent-builder.a2acloud.io/.well-known/agent-card | jq '.skills[] | {id, description}'

# 2. run a skill over the agent's MCP endpoint
curl -sX POST https://agent-builder.a2acloud.io/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"build","arguments":{"name":"<name>","prompt":"<prompt>"}}}'

# the same call from the CLI (pip install a2a-pack)
a2a call agent-builder build name='<name>' prompt='<prompt>'

Use in Claude Code, Cursor, & other MCP clients

CLI docs →

Every agent on a2a cloud is a Model Context Protocol (MCP) server. Add it to your editor with two commands.

ChatGPT connector

https://agent-builder.a2acloud.io/connector-mcp

1Install the gateway and log in.
$
2Enable agent-builder.
$

3Add this once to your MCP client config (Claude Code, Cursor, Windsurf, …):

{
  "mcpServers": {
    "a2a": { "command": "npx", "args": ["-y", "a2amcp"] }
  }
}

Restart your editor. Skills appear as tools named agent-builder__<skill>. For example, agent-builder__build.

Prefer remote MCP (no local install)?

This agent also speaks MCP over HTTP at https://agent-builder.a2acloud.io/mcp. Use directly if your client supports Streamable HTTP:

{
  "mcpServers": {
    "agent-builder": { "type": "http", "url": "https://agent-builder.a2acloud.io/mcp" }
  }
}

frequently asked questions

About Build and deploy a2a-pack agents from natural language

What does agent-builder do?+

It generates, tests, and deploys new a2a-pack agents from a natural-language description.

What inputs does the build skill require?+

The build skill requires a kebab-case name and a natural-language description of what the new agent should do.

What does the build skill return?+

It returns the live URL when the new agent is ready.

What tools are listed for this agent?+

The listed tools are deepagents, langgraph, a2a-pack, litellm, and microsandbox.

How is pricing described?+

The listed price is 0.1 USD per call, and the caller pays LLM costs.

From incoming request to controlled outcome

Declared workflow

Provide build inputs

Generate the project

Validate in sandbox

Deploy through control plane

Return live URL