feat(outbound): A2A server + Vercel deploy for Bio marketplace

[bioanywhere]

What this adds

A complete local-tunnel architecture so the Outbound agent can accept calls from the Bio marketplace and any A2A-compatible caller — while still running on the user's machine with full access to their locally configured Composio integrations (Apify, Airtable, Apollo, Instantly).

New files

store/agents/outbound/server/

File	Purpose
index.js	Express server + full Claude agent loop. Exposes bash, read_file, write_file tools so Claude can follow SKILL.md procedures and call Composio CLI exactly as it does interactively in Houston.
launch.ps1	One-click launcher: kills port 3001, starts the server, starts a Cloudflare Tunnel, waits for the trycloudflare.com URL, then updates the Vercel env var and redeploys — all automatically.
start.bat	Double-click entry point for Windows (calls launch.ps1).
update-tunnel.js	Updates LOCAL_AGENT_URL on the Vercel project and triggers a redeploy each time the tunnel URL changes.
package.json	Dependencies: express, @anthropic-ai/sdk.
README.md	Full architecture docs, setup guide, A2A protocol reference, troubleshooting.

store/agents/outbound/deploy/

File	Purpose
api/index.js	Vercel serverless forwarder — reads LOCAL_AGENT_URL env var, proxies the A2A POST to the local server. Returns a 503 with a clear message if the tunnel is down.
.well-known/agent-card.json	A2A-spec agent card. Bio reads this for health checks and skill discovery.
index.html	Human-readable landing page at the root URL.
vercel.json	CORS headers + Vercel project config.

Architecture

Bio marketplace / external caller
        │  HTTPS POST (A2A JSON-RPC)
        ▼
Vercel  outbound-agent-iota.vercel.app/api   ← stable public URL
        │  HTTP forward  (LOCAL_AGENT_URL env var)
        ▼
Cloudflare Tunnel  *.trycloudflare.com       ← free, no account needed
        │  localhost:3001
        ▼
Express server  (this folder)
        │  reads CLAUDE.md → Claude agent loop → tool_use
        ▼
Anthropic API  →  bash / Composio CLI / workspace files

Why local tunnel instead of a cloud server

The LinkedIn outreach pipeline takes 30–60 minutes — well beyond Vercel's 300s function timeout. A cloud server would also require re-configuring all Composio integrations server-side. The tunnel approach keeps the agent running on the user's machine with zero credential duplication.

Tested on

• Windows 11, Node 20, Vercel CLI 54.6.x, cloudflared 2026.5.2
• Full end-to-end: Bio → Vercel → tunnel → local server → Anthropic API → response ✓
• Bio marketplace health check: ok ✓

Also synced

Latest CLAUDE.md and all 8 SKILL.md files from the Workshop Bogotá session where this was developed and tested.