descrybe/docs/green-chat-openai.md

Green Chat as OpenAI

Canonical wiring guide: green-chat-llm.md — OpenAI-compatible env (OPENAI_*), discovery, and smoke tests.

This page covers LAN reachability when green-chat is bound to loopback only.

---

Current workstation probe (2026-08-04)

Check	Result
Host	192.168.50.181 (ping OK)
Port	8767
Test-NetConnection 192.168.50.181 -Port 8767	TcpTestSucceeded = False
curl http://192.168.50.181:8767/v1/models	Connection timed out
Local http://127.0.0.1:8767	Something else answers (Windows ge chat serve / local GGUF) — not the green host

Cause (expected): green-chat on the green machine was started with --host 127.0.0.1 --port 8767, so it is not reachable over the LAN.

Descrybe root .env (or /integrations/ai) can still hold:

OPENAI_BASE_URL=http://192.168.50.181:8767/v1
OPENAI_MODEL=overloaded-local
OPENAI_API_KEY=…   # set locally; do not commit

…but API/worker calls will fail until bind or tunnel is fixed. Restart api and worker after any process-env change (dashboard AI keys do not require restart).

---

Fix A — Rebind green-chat to 0.0.0.0 (preferred)

On the green machine (SSH as a user with sudo). Inspect the current unit first:

sudo systemctl cat green-chat
sudo systemctl cat green-chat.service.d/*.conf 2>/dev/null
ss -tlnp | grep -E '8767|ge|green|chat'

Create a drop-in that forces LAN bind (adjust ExecStart= to match the real binary/flags from systemctl cat; only change --host):

sudo mkdir -p /etc/systemd/system/green-chat.service.d
sudo tee /etc/systemd/system/green-chat.service.d/10-listen.conf >/dev/null <<'EOF'
[Service]
# Clear inherited ExecStart, then set full command with --host 0.0.0.0
ExecStart=
ExecStart=/usr/local/bin/ge chat serve --host 0.0.0.0 --port 8767
# If your unit uses Environment instead of flags, prefer e.g.:
# Environment=GE_CHAT_HOST=0.0.0.0
# Environment=GE_CHAT_PORT=8767
EOF
sudo systemctl daemon-reload
sudo systemctl restart green-chat
ss -tlnp | grep 8767

Confirm you see 0.0.0.0:8767 (or *:8767), not only 127.0.0.1:8767. Open the firewall if needed:

sudo ufw allow from 192.168.50.0/24 to any port 8767 proto tcp

Then from Windows:

Test-NetConnection 192.168.50.181 -Port 8767
curl.exe -sS -H "Authorization: Bearer YOUR_KEY" http://192.168.50.181:8767/v1/models

Keep OPENAI_BASE_URL=http://192.168.50.181:8767/v1 in root .env or /integrations/ai, restart API + worker if using process env.

---

Fix B — SSH tunnel from Windows (no systemd change)

If you cannot rebind yet, forward loopback from the green host to this PC:

# On Windows (leave this session open)
ssh -L 8767:127.0.0.1:8767 green@192.168.50.181

Then point Descrybe at the tunnel:

OPENAI_BASE_URL=http://127.0.0.1:8767/v1
OPENAI_MODEL=overloaded-local

Port conflict: this Laragon PC already has a local green-chat (or similar) on 127.0.0.1:8767. Stop that local process first, or use a different local port:

ssh -L 18767:127.0.0.1:8767 green@192.168.50.181
# then OPENAI_BASE_URL=http://127.0.0.1:18767/v1

Restart API + worker after changing .env.

---

Related

• green-chat-llm.md — full OpenAI wiring, discovery, app-level smoke checks
• ops-runtime.md — worker / credits