oneport-debug-local 0.2.0

Air-gapped AI debugger for banking, healthcare, and defense: runs entirely on-premise via Ollama/vLLM/llama.cpp with network enforcement to guarantee zero data exfiltration

oneport-debug-local

Air-gapped AI debugger for regulated industries. Banks, hospitals, and defense shops forbid sending any code or logs to external APIs — so Claude Code and ChatGPT are blocked at the firewall, and developers get no AI debugging. oneport-local runs a debugging agent entirely on-premise against your own Ollama / vLLM / llama.cpp model, enforces the air-gap with iptables, and writes a tamper-evident audit trail that proves zero data left the network.

Try it in 60 seconds (no GPU, no model, no API key)

pip install oneport-debug-local
oneport-local demo

The demo runs the real debugging agent against a materialized buggy repo (driven by a bundled offline model), shows the real iptables rules that would seal the network, and writes + verifies a real hash-chained audit entry:

Air-gap compliance preflight
  iptables -A OUTPUT -j DROP
  iptables -I OUTPUT -d localhost -p tcp --dport 11434 -j ACCEPT
  ...
✔ provider=local-inference  ·  hash chain intact

Use it on your own code + on-prem model

export ONEPORT_MODE=local
export LOCAL_INFERENCE_URL=http://localhost:11434       # Ollama, vLLM, or llama.cpp
export LOCAL_INFERENCE_MODEL=deepseek-coder:6.7b

oneport-local debug src/payment/charge.py --error-log /var/log/app.log

The agent reads local files, greps, and runs tests in the workspace — talking only to your local model. It never makes any other network call.

Seal the network (provable air-gap)

sudo -E oneport-local debug src/ --error-log app.log --enforce-network

--enforce-network applies iptables OUTPUT rules that DROP all outbound traffic except the allowlisted local inference endpoint, then removes them when done.

Honesty by design: enforcement requires Linux + root. On any other host the tool records enforcement_status=skipped:* and prints a warning — it will never write network_enforced=true for an air-gap it didn't actually seal. The audit log always reflects what really happened.

Prove compliance

Every analysis writes a hash-chained entry to the audit log. Verify it anytime:

oneport-local verify-compliance --log /var/log/oneport-debug/audit.jsonl

This checks (a) every session used local inference only — any cloud provider is a violation — and (b) the hash chain is intact, so any edit, deletion, or insertion of a past entry is detected. Required for SOX §404, HIPAA §164.312(b), PCI-DSS 10.5, ISO 27001 A.12.4.

oneport-local models --server http://localhost:11434   # list available local models

Configuration

Variable	Meaning
ONEPORT_MODE=local	Forces local inference; never falls back to cloud
LOCAL_INFERENCE_URL	Ollama/vLLM/llama.cpp endpoint (must be in the host allowlist)
LOCAL_INFERENCE_MODEL	Model name, e.g. deepseek-coder:6.7b
AUDIT_LOG_PATH	Where the hash-chained audit trail is written
AUDIT_HMAC_KEY	Signing key for the audit chain (set this in production)

The local-inference provider refuses to connect to any non-allowlisted host — a misconfigured URL fails closed rather than silently reaching the internet.

Why an agent, not just a prompt

The agent guards its own sandbox: file reads are confined to the workspace (path traversal is blocked), and destructive commands (rm, git push, curl, pip install, …) are refused. It's built to run unattended inside a regulated network.

License

Apache-2.0