a2a-pack 0.1.13

Developer SDK + CLI for building, packaging, and deploying A2A agents.

a2a-pack

Developer SDK + CLI for building, packaging, and deploying full A2A compliant agents.

One Python class becomes a sandboxed, discoverable, MCP-compatible, full A2A compliant AI agent on the a2a cloud platform. Other agents reach yours via HMAC-signed grants. The platform owns deployment, execution, permissions, and (when you're ready) billing.

a2a-pack makes Google's Agent2Agent protocol easy: Agent Cards, tasks, messages, artifacts, file/data exchange, streaming, auth, JSON-RPC, REST, and protocol errors are handled by the SDK runtime.

pip install a2a-pack
a2a signup --email [e-mail] --password ...
a2a init research-agent
cd research-agent
a2a dev
a2a test --invoke
a2a deploy
# -> https://research-agent.a2acloud.io   (A2A, MCP, TLS, all wired)

What an agent looks like

from pydantic import BaseModel
from a2a_pack import (
    A2AAgent, LLMProvisioning, NoAuth, Pricing, RunContext, skill,
)


class GreeterConfig(BaseModel):
    suffix: str = "!"


class Greeter(A2AAgent[GreeterConfig, NoAuth]):
    name = "greeter"
    description = "Say hi."
    version = "0.1.0"
    config_model = GreeterConfig
    auth_model = NoAuth

    # Use the caller's own LLM key (forwarded by the platform) — the
    # author's price stays small; the LLM bill goes to the caller's
    # provider directly.
    llm_provisioning = LLMProvisioning.CALLER_PROVIDED
    pricing = Pricing(price_per_call_usd=0.01, caller_pays_llm=True)

    @skill(description="Greet someone.")
    async def greet(self, ctx: RunContext[NoAuth], who: str) -> str:
        await ctx.emit_progress(f"greeting {who}")
        return f"hello {who}{self.config.suffix}"

That's it. a2a deploy packages the source, the control plane builds the image, ArgoCD reconciles, you get a full A2A compliant public URL.

Local development

Use the same agent card, invoke path, secret names, and workspace contract before uploading anything:

a2a dev

That starts the agent at http://127.0.0.1:8000, loads .env.local, creates .a2a/workspace/{inputs,outputs}, and enables hot reload. Skills are callable at POST /invoke/{skill} and the card is visible at /.well-known/agent-card.

It also serves a bundled local console at http://127.0.0.1:8000/_dev with drag-and-drop file upload into inputs/, workspace file previews/downloads, schema-derived skill inputs, streaming progress, and final results.

Run preflight checks before deploy:

a2a test
a2a test --invoke --skill summarize --args-json '{"text":"hello"}'

Secrets stay local in .env.local. Workspace-backed framework tools, including DeepAgents via ctx.workspace_backend(), write durable local outputs under .a2a/workspace/outputs so you can inspect what will become downloadable files in A2A Cloud.

If your skill runs code that creates files, use ctx.workspace_shell(...) or ctx.workspace_python(...). Those commands run in the platform sandbox with the caller's workspace mounted, so /workspace/... writes persist directly and changed files elsewhere in the sandbox root filesystem are mirrored under outputs/rootfs-captures/.... Plain in-process subprocess calls inside the agent container are not workspace-mounted or rootfs-captured.

Bring your own auth

Agents can make caller identity explicit by declaring an auth_model and an auth_resolver. The resolver receives the inbound bearer token, validates it against your auth system, and returns the typed principal skills read from ctx.auth.

from a2a_pack import A2AAgent, JWTAuth, OIDCUserInfoAuthResolver, RunContext, skill


class CustomerAgent(A2AAgent):
    name = "customer-agent"
    description = "Uses the caller's app login"
    auth_model = JWTAuth
    auth_resolver = OIDCUserInfoAuthResolver(
        "https://auth.example.com/oauth2/userinfo",
        auth_model=JWTAuth,
    )

    @skill(scopes=["profile:read"])
    async def profile(self, ctx: RunContext[JWTAuth]) -> dict[str, str | None]:
        return {"sub": ctx.auth.sub, "email": ctx.auth.email}

For custom APIs, subclass AuthResolver and call your own /me, /introspect, or session-exchange endpoint. SAML-backed apps use the same contract by exposing a bearer-token bridge endpoint that returns JSON.

Public surface

Concept	Where
A2AAgent base class + @skill decorator	a2a_pack.agent
RunContext, ctx.llm, ctx.ask, ctx.request_scope	a2a_pack.context
Grant mint/verify (HMAC, audience-bound, glob-filtered, time-limited)	a2a_pack.grants
Workspace negotiation surface	a2a_pack.workspace
Sandbox client (microVM via libkrun)	a2a_pack.sandbox
Agent-to-agent client (HTTP, in-memory, custom)	a2a_pack.a2a_client
MCP server (skills → tools, mountable into your FastAPI app)	a2a_pack.mcp
Lifecycle / Resources / Pricing / LLMProvisioning declarations	a2a_pack.runtime
Card schema (auto-derived from your class)	a2a_pack.card

A2A made easy

• Subclass A2AAgent.
• Decorate async methods with @skill.
• Use RunContext for progress, auth/input requests, artifacts, scoped files, sandbox execution, and agent-to-agent calls.
• Run a2a deploy; the runtime exposes full A2A plus MCP and HTTPS.

Full reference + auto-generated docs at https://docs.a2acloud.io.

Self-hosting

The platform pieces (control plane, sandbox runtime, gitea, ArgoCD, MinIO, LiteLLM) live at gitea.a2acloud.io — the SDK is the only piece you need on PyPI. If you want to run the whole stack locally or in your own cluster, the bootstrap recipe is in the platform README.

License

MIT — see LICENSE.