Claude Code

Desktop

The standalone Desktop app connects via Settings → Connection → Gateway:

Connection:          Gateway

Gateway base URL:    https://api.opengateway.one/oss

Gateway API key:     YOUR_KEY

Gateway auth scheme: bearer

Model list:          leave empty

Click Test connection. With model discovery enabled, each picker entry is labeled From gateway and shows the real open model via display_name — e.g. GLM 5.2, Kimi K2.7 Code, MiniMax M3. Wire ids stay claude-* for Desktop compatibility; labels are never Sonnet/Opus/Haiku tiers.

For 1M context on GLM 5.2 or MiniMax M3, Desktop does not expose custom headers — set CLAUDE_CODE_MAX_CONTEXT_TOKENS and DISABLE_COMPACT in ~/.claude/settings.json under env, then quit and relaunch Desktop (see troubleshooting).

IDE (VS Code & JetBrains)

The VS Code and JetBrains extensions share configuration with the CLI via ~/.claude/settings.json. Set the gateway in the env block:

{

  "env": {

    "ANTHROPIC_BASE_URL": "https://api.opengateway.one/oss",

    "ANTHROPIC_AUTH_TOKEN": "YOUR_KEY",

    "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1"

  }

}

Reload the window after changing settings. The IDE picker uses the same Anthropic-shaped discovery as Desktop — entries show real open-weight names including GLM 5.2, Kimi K2.x, and Qwen3 Coder. See Model discovery for the usable-model contract.

CLI

Set the base URL and auth via environment variables (or ~/.claude/settings.json under env):

export ANTHROPIC_BASE_URL="https://api.opengateway.one/oss"

export ANTHROPIC_AUTH_TOKEN="YOUR_KEY"   # → Authorization: Bearer

# or: export ANTHROPIC_API_KEY="YOUR_KEY"  # → x-api-key

Pass the model with friendly open-model ids:

claude --model glm-5.2 "Refactor the auth module."

claude --model kimi-k2.6 "Add tests for the payment handler."

claude --model qwen3-coder "Implement the retry logic."

Raw catalog ids with routing policies also work:

claude --model "zai-org/GLM-5.2:preferred" "Profile this hot path."

claude --model "moonshotai/Kimi-K2.7-Code:preferred" "Review the API layer."

Model discovery (optional)

Claude Code only calls the gateway's models endpoint when you opt in:

export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1

The picker filters to ids beginning claude/anthropic (a Claude Code client-side constraint). OpenGateway sets display_name to the native open model — e.g. GLM 5.2 — so IDE and Desktop pickers stay honest. Discovery is optional; CLI ids like glm-5.2 work directly through /v1/messages. See Model discovery.

1M context (GLM 5.2, MiniMax M3)

OpenGateway advertises the real context window per model. GLM 5.2 and MiniMax M3 expose context_window: 1000000 and metadata.context_tier: "one-million" in discovery. No beta header is required on the OSS lane.

{

  "env": {

    "ANTHROPIC_BASE_URL": "https://api.opengateway.one/oss",

    "ANTHROPIC_AUTH_TOKEN": "YOUR_KEY",

    "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1",

    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "1000000",

    "DISABLE_COMPACT": "1"

  },

  "model": "claude-sonnet-4-6"

}

Set both CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000 and DISABLE_COMPACT=1 — required on current Desktop builds when the discovery wire id is a known claude-* alias. Without DISABLE_COMPACT=1, Claude Code ignores MAX_CONTEXT_TOKENS per official env-var docs. No custom headers or beta headers are needed on the OSS lane.

Pick GLM 5.2 or MiniMax M3 from the gateway picker, or pass claude --model glm-5.2 on CLI. See 1M context.

If the context panel shows 200K instead of 1M

Symptom: Desktop shows GLM 5.2 from the gateway picker, but the context panel reads 200.0k (200K).

Why: Discovery keeps wire ids like claude-sonnet-4-6 (a Claude Code Desktop requirement). Labels come from display_name, but Desktop still defaults known claude-* ids to Anthropic's 200K unless you set CLAUDE_CODE_MAX_CONTEXT_TOKENS client-side. OpenGateway already returns context_window: 1000000 and metadata.context_tier: "one-million".

1. Add to ~/.claude/settings.json under env: CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1, CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000, and DISABLE_COMPACT=1. Both context vars are required — DISABLE_COMPACT=1 is not optional; Claude Code ignores CLAUDE_CODE_MAX_CONTEXT_TOKENS unless compaction is disabled.
2. Desktop has no custom-headers UI — env vars in settings.json are the only client-side override. No anthropic-beta header is needed on OSS.
3. Quit and relaunch Claude Code Desktop (not just reload a window) so env vars and the discovery cache reload.
4. CLI bypass: claude --model glm-5.2 uses the native id and avoids the wire-id cap.

Verify the gateway: curl -s https://api.opengateway.one/oss/v1/models — GLM 5.2 entries should show all three context fields at 1000000. No beta header is required on OSS.

Effort

Extended thinking is on by default. Newer models use CLAUDE_CODE_EFFORT_LEVEL (low/medium/high); OpenGateway maps it onto whatever the upstream model expects — see Effort levels.

Models to try

Call GET /oss/v1/models for the full live catalog. Lane-backed discovery aliases (e.g. claude-sonnet-4-6 → GLM 5.2) exist for picker compatibility; CLI and scripted use should prefer the CLI id column.