API Usage Guide

Connection

Base URL	`https://api.nghimmo.com`
Key	sk-xxxxxxxxxxxx

No /v1. Cursor Pro is the exception — add /v1.

Models

Strongest, hard tasks. Latest: nghi/claude-opus-5.

nghi/claude-opus-5
nghi/claude-opus-5-thinking
nghi/claude-opus-4.8
nghi/claude-opus-4.8-thinking
nghi/claude-opus-4.7
nghi/claude-opus-4.7-thinking
nghi/claude-opus-4.6
nghi/claude-opus-4.6-thinking
nghi/claude-opus-4.5
nghi/claude-opus-4.5-thinking

Daily use. Default: nghi/claude-sonnet-5.

nghi/claude-sonnet-5
nghi/claude-sonnet-4.6
nghi/claude-sonnet-4.6-thinking
nghi/claude-sonnet-4.5
nghi/claude-sonnet-4.5-thinking
nghi/claude-sonnet-4
nghi/claude-sonnet-4-thinking

Light work, cheaper tokens.

nghi/claude-haiku-4.5
nghi/claude-haiku-4.5-thinking

Server picks a model.

nghi/auto

Setup

Install Claude Desktop: claude.com/download — not Microsoft Store.
Download the tool, paste key → Apply:
macOS Windows Linux
Reopen the app → Continue with Gateway.

If your OS blocks the tool: macOS right-click → Open · Windows → More info → Run anyway. On Linux use the Linux tab (Claude Desktop rarely ships for Linux).

New key: paste → Apply. Remove: Reset.

Mac — tool OK but Gateway not showing

Install Claude from the .dmg at claude.com/download — not the App Store.
Cmd+Q to fully quit Claude before running the tool.
If double-click fails, use Terminal:

cd ~/Downloads
xattr -d com.apple.quarantine ClaudeKeyTool.command 2>/dev/null
chmod +x ClaudeKeyTool.command
bash ClaudeKeyTool.command

Verify config was written:

ls ~/Library/Application\ Support/Claude-3p/configLibrary/
cat ~/Library/Application\ Support/Claude-3p/configLibrary/_meta.json

Open Claude → choose Continue with Gateway (do not sign in to claude.ai). Tool log: ~/Library/Logs/claude-key-tool.log

npm install -g @anthropic-ai/claude-code
Run ClaudeKeyTool (same tool — configures Claude Code too):
macOS Windows Linux

Manual setup

Edit ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.nghimmo.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-xxxxxxxxxxxx",
    "ANTHROPIC_MODEL": "nghi/claude-sonnet-5[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "nghi/claude-opus-5[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "nghi/claude-sonnet-5[1m]",
    "ANTHROPIC_SMALL_FAST_MODEL": "nghi/claude-haiku-4.5[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "nghi/claude-haiku-4.5[1m]",
    "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1",
    "CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK": "1",
    "API_TIMEOUT_MS": "600000"
  }
}

Leave ANTHROPIC_API_KEY empty — use ANTHROPIC_AUTH_TOKEN only. 10-minute timeout (API_TIMEOUT_MS=600000) so long thinking is not cut off. GATEWAY_MODEL_DISCOVERY loads models from the gateway. DISABLE_REFUSAL_FALLBACK prevents falling back to Anthropic.com.

Claude Code CLI does not read context from the server — it defaults to /200k. You must append [1m] to the model name (e.g. nghi/claude-opus-5[1m]) or run /model nghi/claude-opus-5[1m], then restart CLI. Backend already accepts 1M.

On Linux prefer Claude Code CLI or the VS Code / Cursor extension. Official Claude Desktop usually has no Linux build.

Install Node.js 18+, then:

npm install -g @anthropic-ai/claude-code

Download the Linux script and run it in Terminal:

Linux (.sh)

chmod +x ClaudeKeyTool.sh
./ClaudeKeyTool.sh sk-xxxxxxxxxxxx

Run claude. To switch models: /model nghi/claude-sonnet-5[1m]

VS Code / Cursor / Codium on Linux: install the Claude Code extension → run the same script → Ctrl+Shift+P → Reload Window.

Manual setup (no script)

Edit ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.nghimmo.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-xxxxxxxxxxxx",
    "ANTHROPIC_MODEL": "nghi/claude-sonnet-5[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "nghi/claude-opus-5[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "nghi/claude-sonnet-5[1m]",
    "ANTHROPIC_SMALL_FAST_MODEL": "nghi/claude-haiku-4.5[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "nghi/claude-haiku-4.5[1m]",
    "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1",
    "CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK": "1",
    "API_TIMEOUT_MS": "600000"
  }
}

Quick curl test

curl https://api.nghimmo.com/v1/messages \
  -H "x-api-key: sk-xxxxxxxxxxxx" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"nghi/claude-sonnet-5","max_tokens":64,"messages":[{"role":"user","content":"ping"}]}'

Remove config: ./ClaudeKeyTool.sh --reset

Install the Claude Code extension in VS Code / Cursor / Antigravity.
Run ClaudeKeyTool → paste key → Apply:
macOS Windows Linux
Ctrl+Shift+P → Reload Window.

Ctrl+Shift+J → Models
Enable Override OpenAI Base URL → https://api.nghimmo.com/v1
Paste key → Verify → add nghi/claude-sonnet-5

Turn off all of Cursor's default models.

Chatbox → Settings → Model Provider → Claude
API Host https://api.nghimmo.com/v1 · Key sk-...
Add nghi/claude-sonnet-5 → Test Model

Connection

Base URL	`https://api.nghimmo.com/v1`
Key	sk-xxxxxxxxxxxx

Always with /v1. Grok keys don't run Claude Code.

Models

Send no model and you get nghi/grok-4.5.

Model	Use for
nghi/grok-4.5	Strongest — code, agents
nghi/grok-4.3	1M context, cheaper
nghi/grok-4.20-0309-reasoning	With reasoning
nghi/grok-4.20-0309-non-reasoning	Fast, cheap
nghi/grok-4.20-multi-agent-0309	Deep research
nghi/grok-build-0.1	Code / agent
nghi/grok-composer-2.5-fast	Fast coding in Cursor
nghi/grok-4.5-latest nghi/grok-build-latest	Aliases of grok-4.5
nghi/grok-3-mini nghi/grok-3-mini-fast	Legacy

These 11 only. A wrong name errors out. No image/video models.

Setup

Ctrl+Shift+J → Models
Enable Override OpenAI Base URL → https://api.nghimmo.com/v1
Paste key → Verify → add nghi/grok-4.5

Turn off all of Cursor's default models.

Cline / Roo Code: Settings → API Provider OpenAI Compatible → fill in:

Base URL : https://api.nghimmo.com/v1
API Key  : sk-xxxxxxxxxxxx
Model    : nghi/grok-4.5

Continue: edit ~/.continue/config.yaml:

models:
  - name: Nghimmo Grok
    provider: openai
    model: nghi/grok-4.5
    apiBase: https://api.nghimmo.com/v1
    apiKey: sk-xxxxxxxxxxxx

LobeChat, NextChat, Open WebUI… — any OpenAI-compatible app:

Base URL : https://api.nghimmo.com/v1
API Key  : sk-xxxxxxxxxxxx
Model    : nghi/grok-4.5

Chatbox → Settings → Model Provider → OpenAI API Compatible
API Host https://api.nghimmo.com/v1 · Key sk-...
Add nghi/grok-4.5 → Test

Connection

Base URL	`https://api.nghimmo.com/v1`
Key	sk-xxxxxxxxxxxx

Always with /v1. Codex keys don't run Claude Code.

Models

Model	Use for
nghi/gpt-5.5	Default — stable
nghi/gpt-5.6-sol	Strongest, token-hungry
nghi/gpt-5.6-terra	Balanced
nghi/gpt-5.6-luna	Fast, cheap
nghi/gpt-5.4	Large context
nghi/gpt-5.4-mini	Small tasks, subagents
nghi/codex-auto-review	Automatic code review

Codex keys reject Claude and Grok models.

Setup

Install ChatGPT from Microsoft Store (by OpenAI).
Download the tool → Desktop → paste key → Sign in:
macOS Windows Linux
Quit ChatGPT fully, reopen → use ChatGPT Work / Codex.

Remove: tool → Reset. On Linux use the Linux tab (Codex CLI).

On Linux use Codex CLI or the Codex VS Code extension. ChatGPT Desktop usually has no Linux build.

Install Codex CLI (per OpenAI / npm docs for your CLI build).

Download the script and run:

Linux (.sh)

chmod +x CodexKeyTool.sh
./CodexKeyTool.sh sk-xxxxxxxxxxxx
codex

Codex extension (VS Code) on Linux: install the extension → run the script above → fully quit VS Code and reopen.

Manual setup (~/.codex)

~/.codex/auth.json:

{
  "auth_mode": "apikey",
  "OPENAI_API_KEY": "sk-xxxxxxxxxxxx"
}

~/.codex/config.toml (short form):

model = "gpt-5.6-sol"
model_provider = "Nghimmo"
sandbox_mode = "danger-full-access"
approval_policy = "never"

[model_providers.Nghimmo]
name = "Nghimmo"
base_url = "https://api.nghimmo.com/v1"
env_key = "OPENAI_API_KEY"
wire_api = "responses"

Quick curl test

curl https://api.nghimmo.com/v1/chat/completions \
  -H "Authorization: Bearer sk-xxxxxxxxxxxx" \
  -H "content-type: application/json" \
  -d '{"model":"nghi/gpt-5.5","messages":[{"role":"user","content":"ping"}]}'

Remove: ./CodexKeyTool.sh --reset

Ctrl+Shift+J → Models
Enable Override OpenAI Base URL → https://api.nghimmo.com/v1
Paste key → Verify → add nghi/gpt-5.5

Turn off all of Cursor's default models.

Codex extension (OpenAI) in VS Code:

Install the Codex extension (OpenAI) in VS Code.
Download the tool → choose CLI → paste key → Sign in:
Windows macOS Linux
Fully quit VS Code → reopen → start a new Codex chat.

Update Agent sandbox on Windows: open %USERPROFILE%\.codex\config.toml, ensure sandbox = "unelevated" under [windows], then re-run the tool or edit manually.

Cline / Roo Code: Settings → API Provider OpenAI Compatible → fill in:

Base URL : https://api.nghimmo.com/v1
API Key  : sk-xxxxxxxxxxxx
Model    : nghi/gpt-5.5

Continue: edit ~/.continue/config.yaml:

models:
  - name: Nghimmo Codex
    provider: openai
    model: nghi/gpt-5.5
    apiBase: https://api.nghimmo.com/v1
    apiKey: sk-xxxxxxxxxxxx

LobeChat, NextChat, Open WebUI… — any OpenAI-compatible app:

Base URL : https://api.nghimmo.com/v1
API Key  : sk-xxxxxxxxxxxx
Model    : nghi/gpt-5.5

Chatbox → Settings → Model Provider → OpenAI API Compatible
API Host https://api.nghimmo.com/v1 · Key sk-...
Add nghi/gpt-5.5 → Test

Connection

Base URL	`https://api.nghimmo.com/v1`
Key	Cursor API Key (create in admin → API Cursor tab)

Use a Cursor key (created in admin), not a Claude/Grok/Codex key. Routes through Cursor OAuth — Grok Bot weekly quota applies.
Note: composer-2.5 is Cursor's agent model and only works well inside Cursor IDE; other clients (VS Code/Copilot/Cline) doing code tasks often 400 → prefer nghi/grok-4.6.

Models

Composer

Model	Note
nghi/composer-2.5	Default Composer (Cursor IDE only)
nghi/composer-2.5-fast	Fast Composer (Cursor IDE only)

Grok (23 models)

Grok 4.6 is the latest in Cursor. Default: nghi/grok-4.6. Recommended for non-Cursor clients.

Model	Note
nghi/grok-4.6	Latest — Cursor Grok 4.6 (recommended)
nghi/grok-4.6-fast	Grok 4.6 Fast
nghi/grok-4.5	Grok 4.5
nghi/grok-4.5-high	Grok 4.5 High
nghi/grok-4.5-medium	Grok 4.5 Medium
nghi/grok-4.5-low	Grok 4.5 Low
nghi/grok-4.3	Grok 4.3
nghi/grok-4.20	Grok 4.20
nghi/grok-4.2	Grok 4.2 (4.20 Beta)
nghi/grok-4.1-thinking	Grok 4.1 Thinking
nghi/grok-4.1-expert	Grok 4.1 Expert
nghi/grok-4.1-fast	Grok 4.1 Fast
nghi/grok-4.1-mini	Grok 4.1 Mini (Thinking)
nghi/grok-4-heavy	Grok 4 Heavy (SuperGrok)
nghi/grok-4-thinking	Grok 4 Thinking
nghi/grok-4-fast-reasoning	Grok 4 Fast Reasoning
nghi/grok-4-mini	Grok 4 Mini (Thinking)
nghi/grok-4	Grok 4
nghi/grok-3-thinking	Grok 3 Thinking
nghi/grok-3-mini	Grok 3 Mini (Thinking)
nghi/grok-3	Grok 3
nghi/grok-code-fast-1	Grok Code Fast 1

All 25 models route through the Cursor account. Incorrect model name = error. For code tasks on non-Cursor clients, prefer Grok, avoid Composer.

Setup

Cline / Roo Code: Settings → API Provider OpenAI Compatible:

Base URL : https://api.nghimmo.com/v1
API Key  : (Cursor key from admin)
Model    : nghi/grok-4.6

Continue: edit ~/.continue/config.yaml:

models:
  - name: Cursor Grok 4.6
    provider: openai
    model: nghi/grok-4.6
    apiBase: https://api.nghimmo.com/v1
    apiKey: (Cursor key from admin)

Composer (nghi/composer-2.5) only works well inside Cursor IDE. Other clients should use Grok.

cURL / Python / any OpenAI SDK:

curl https://api.nghimmo.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_CURSOR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"nghi/grok-4.6","messages":[{"role":"user","content":"Hello"}]}'

API Usage Guide

Connection

Models

Setup

Connection

Models

Setup

Connection

Models

Setup

Connection

Models

Composer

Grok (23 models)

Setup

Save Tokens

Tokens Left