Skip to main content

Gateway integration

The gpt2giga gateway accepts requests in common client formats and sends them to GigaChat. The GigaLoom base package is independent; install the integration only for GigaChat routes or Direct Chat.

uv tool install --force \
  --with 'gpt2giga>=0.3.0,<0.4.0' \
  'gigaloom[gpt2giga]==0.9.1'

GigaLoom uses the published package. A sibling gpt2giga source checkout is not required.

Select a model and route​

The convenience form selects a gateway and model:

giga --with gpt2giga --model GigaChat-2-Max codex

Resolve the choice without launching an agent or sending a model request:

giga --with gpt2giga --model GigaChat-2-Max --dry-run --json codex

You can also select an immutable route id:

giga --route codex-gpt2giga-gigachat-2-max codex --help

Global GigaLoom options must precede the agent name. Therefore giga --model GigaChat-2-Max codex selects a route, while giga codex --model o3 passes --model o3 to Codex.

The model list comes from /models. GigaLoom does not infer capabilities from model names or substitute a different model when the selection is unavailable.

Protocols and agents​

gpt2giga exposes several public APIs, but an agent can support only one of them:

Agent or clientProtocol to the gatewayStatus
CodexOpenAI ResponsesTechnical preview; inspect the dry-run result
OpenCode and compatible ACP agentsOpenAI Chat CompletionsAvailable when inspect reports provider_bridge.status=ready
Claude CodeAnthropic MessagesRequires a separately verified client route
Gemini CLIGemini GenerateContentA custom endpoint can be unavailable for the current client version
Amp or an unknown ACP agentNative model providerNo gateway is selected without a verified configuration method

native_only is a capability boundary, not a failed installation. Launch such an agent without --with and --model.

Launch an ACP agent through the gateway​

Inspect the agent after Registry installation:

giga agent inspect opencode --json
giga agent probe opencode --json

When the state is ready, GigaLoom creates temporary configuration, starts the ACP process, applies the verified model-provider address, and only then creates a session:

giga --with gpt2giga --model GigaChat-2-Max opencode

Temporary configuration disappears with the process. GigaLoom does not write to ~/.opencode, ~/.codex, ~/.qwen, ~/.claude, or ~/.gemini.

If an agent advertises configurable providers but rejects the update or does not confirm the readback, GigaLoom blocks launch. It does not fall back to unverified environment variables.

Managed and external gateways​

A managed profile starts the installed gpt2giga, waits for readiness, and reuses a healthy process. On cleanup, GigaLoom stops only a process it owns.

An external profile connects to a gateway that you already run. You own its lifecycle; GigaLoom checks the address, authentication, model list, and API compatibility.

In both modes, secrets are resolved at launch and do not appear in inspect, stored probe results, or route URLs.

If a route is unavailable​

GigaLoom stops launch when it cannot confirm compatibility. It does not change the protocol, gateway, model provider, model, or agent without your choice.

Check that:

  1. the optional gpt2giga integration is installed;
  2. GigaChat credentials are available in the current terminal;
  3. /models returns the selected model;
  4. giga agent inspect <id> --json reports ready;
  5. another process does not own the managed gateway port.

Match reason_ids with Launch troubleshooting. Native launch without a gateway remains available when only the route is affected.

gpt2giga reference​