zipcode Cloud Relay: Field Coding Agent for Low-Connectivity Teams

zipcode Cloud Relay is a local-first coding agent prototype for field teams working in air-gapped, low-connectivity, and privacy-sensitive environments. The Google Cloud Rapid Agent Hackathon direction extends the local agent into a hybrid workflow: local evidence gathering first, then a human-approved relay to Gemini, Google Cloud Agent Builder, and GitLab MCP when connectivity returns.

The prototype is intentionally small enough to audit:

• Python stdlib only
• local workspace sandboxing
• local search and file-reading tools
• Ollama-compatible local model backend in the offline prototype
• Google Cloud Agent Builder handoff payload for the connected relay
• GitLab MCP workflow action plan for issues, merge requests, and review notes
• deterministic dry-run mode for judges without the model installed

Quick Demo

python zipcode_agent.py `
  --root examples/airgap_demo_workspace `
  --task "Find the bug in the water sensor parser and propose a safe fix." `
  --read README.md `
  --search TODO `
  --dry-run

With a local model running through Ollama:

$env:GEMMA_MODEL="gemma4:e2b"
python zipcode_agent.py `
  --root examples/airgap_demo_workspace `
  --task "Find the bug in the water sensor parser and propose a safe fix." `
  --read README.md `
  --search TODO

Then create the connected relay outbox:

python cloud_relay.py `
  --task "Find the bug in the water sensor parser and propose a safe fix." `
  --report relay_outbox/offline_report.md `
  --approved-by "field operator"

Or run both sides:

powershell -ExecutionPolicy Bypass -File scripts/demo.ps1

Why This Matters

Many useful coding situations are not cloud-friendly: disaster-response laptops, clinics with sensitive patient tooling, rural schools, public-interest NGOs, and teams working under unstable connectivity. In those places, a cloud coding assistant is either unavailable or inappropriate.

zipcode turns local model reasoning into a field coding loop:

1. collect only local evidence,
2. ask the local model for a grounded plan,
3. preserve a transcript for audit,
4. let a human operator approve changes.

Architecture

operator task
  -> workspace sandbox
  -> local tools: list_files, read_file, search_text
  -> local model through Ollama-compatible chat
  -> patch plan + tests + audit transcript
  -> human approval
  -> Google Cloud Agent Builder handoff payload
  -> GitLab MCP action plan

The agent refuses to read outside the selected workspace. It does not need network access after the local model is installed. The Cloud Relay companion adds a connected approval queue so approved evidence bundles can be routed to Gemini and GitLab MCP without uploading an entire repository blindly.

Google Cloud Rapid Agent Track Fit

• Partner track: GitLab.
• Google Cloud fit: Agent Builder + Gemini can orchestrate the connected relay.
• GitLab MCP fit: code-agent output naturally maps to issues, merge requests, review summaries, and audit updates.
• Impact fit: low-connectivity field teams keep sensitive source local while still getting structured software workflow help.

Relationship to Hollow AgentOS

This submission is a focused, judge-friendly slice inspired by the local-agent work in hollow-agentOS. Hollow explores long-running autonomous local agents; zipcode narrows that idea into one clear workflow: local coding help for air-gapped places.

License

MIT