A self-hosted coding agent running on GMI Cloud AgentBox, wired to GMI MaaS for inference. This page watches every session live — status, tool calls, and arguments as they happen — and can drive them too.
Every setting has a working default, so this runs with no environment variables at all. Each one below is an override you can set if you want to change something.
Defaults are baked into the image: a shared trial key and DeepSeek V4 Flash, against GMI MaaS.
model = "deepseek-ai/
DeepSeek-V4-Flash-0731"
base_url = "https://api.
gmi-serving.com/v1"
The chat API identifies models by slug. AgentBox's GMI_MODELS UUID is a different namespace and is ignored — it returns 404 here.
GMI_MODEL="deepseek-ai/…" ✅ GMI_MODELS="7f3d…-uuid" ❌ 404
All optional. Or change the key and model live from the panel below — no restart.
GMI_API_KEY=… own key
GMI_MODEL=… own model
AMCP_USE_AGENTBOX_MODEL_ID=1
use the UUID
AMCP_CHAT_MODEL → GMI_MODEL → default slug (GMI_MODELS only if opted in)GMI_API_KEY → GMI_MAAS_API_KEY → OPENAI_API_KEY → bundled trial key:8080 (public) → amcp serve on 127.0.0.1:8081This deployment ships with a shared trial key so it works immediately. After the free prompts are used, add your own key to keep going — it is written to the running config and takes effect on your next prompt.
The test creates a throwaway session, sends a one-word prompt, and reads the result — the only way to confirm the key and model slug actually resolve. Failures surface as provider errors rather than crashes, so it is safe to re-run. Note that it spends one free prompt while the trial key is in use.
| Tool | Calls | Total ms |
|---|
Counts are observed live from this session's event stream, so they start at zero each time you attach. Bars share one hue because this encodes magnitude, not identity — the tool names on the axis carry identity.
Built-in tools
Agent types
Spin up your own agents on GMI Cloud — bring a key, pick a model, and go from a blank workspace to a running coding agent in minutes.
Open GMI Cloud console