TSG LLM · Public knowledge base

Connect your coding tools

Use one TSG LLM API key with compatible coding assistants, local terminals, or a cloud CLI environment. The public HTTPS endpoint works without a private-network connection.

1. Get your account and API key

1. Ask your administrator for an account, then sign in and replace the temporary password.
2. Create a separate API key for each tool or environment. Copy it when shown; it cannot be displayed again.
3. Store it in your password manager or your cloud platform's secret store. Your account password is not an API key.

Current rates, your output limit, your spending cap, and recorded charges appear in your account. Account access and API credentials are never included in this public guide.

2. Codex CLI

Download the Codex configuration. Back up your existing ~/.codex/config.toml, then merge these settings. Top-level settings go before the first table. If you already specify a local model catalog, remove that conflicting setting for this provider.

Read your key without placing it in shell history. This works in bash and zsh:

printf 'TSG LLM API key: '
read -r -s TSG_LLM_API_KEY
printf '\n'
export TSG_LLM_API_KEY
codex

On Windows, use your secret manager to set TSG_LLM_API_KEY in the current terminal or use a secure prompt:

$secret = Read-Host 'TSG LLM API key' -AsSecureString
$env:TSG_LLM_API_KEY = [System.Net.NetworkCredential]::new('', $secret).Password
codex
Remove-Item Env:TSG_LLM_API_KEY

The configuration uses the Responses API, a model catalog protected by your API key, and local conversation compaction. It starts summarizing long conversations at 190,000 tokens to leave space for replies within the shared 262,144-token context. This is a conservative setup policy; a very large single message can still exceed the context limit. See the Codex configuration reference. Keep the client's sandbox and approval settings appropriate for your project.

3. Claude Code

Download the launcher for your terminal:

• macOS / Linux shell launcher
• Windows PowerShell launcher

Run the launcher from your project directory. It asks for a key if one is not already supplied by your environment:

bash /path/to/tsg-claude-2026-10-23-v2.sh

PowerShell:

.\tsg-claude-2026-10-23-v1.ps1

The default requested output is 32,768 tokens. If your administrator sets a different account limit, set TSG_LLM_MAX_OUTPUT_TOKENS accordingly before launching. The launcher changes settings for that launch; it does not overwrite your global Claude configuration.

The underlying model is Qwen. This does not provide Anthropic's Claude models. Anthropic does not officially support routing Claude Code to non-Claude models. Retest after client upgrades.

4. Other coding assistants

Choose a custom OpenAI-compatible or Anthropic-compatible provider, if the tool offers one. Use these public connection details:

Some clients append /v1 themselves. If requests return 404, check that it appears only once. A subscription-only tool may not permit a custom provider.

5. Cloud CLI environments

The same base URLs work from a cloud terminal with outbound HTTPS access. Add your key through that platform's secret store and expose it as TSG_LLM_API_KEY. Keep the configuration file free of credentials. Install the desired CLI using its official instructions, then use the configuration or launcher above.

Use a separate key for each cloud environment and revoke it when the environment is retired. Review that platform's own handling of source code, command output, and environment variables. Individual cloud products and editor extensions require their own compatibility check.

Check the connection

List the available public model without generating a billable response:

curl --fail --silent --show-error \
  https://llm.tsgglobal.world/v1/models \
  -H "Authorization: Bearer $TSG_LLM_API_KEY"

For a minimal Chat Completions request:

curl --fail --silent --show-error \
  https://llm.tsgglobal.world/v1/chat/completions \
  -H "Authorization: Bearer $TSG_LLM_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen3.8-flash-next-nvfp4","messages":[{"role":"user","content":"Reply with OK."}],"max_tokens":16}'

Limits, billing, and supported features

• Default output limit: 32,768 tokens per request; administrators can configure account limits. The 262,144-token context is shared by the input and response.
• Up to 30 requests per minute and 2 simultaneous model calls per account, within shared service capacity. All keys on an account share these limits.
• Spending caps reserve a conservative maximum before accepting a request. A reservation is not a charge. Final recorded usage determines the charge.
• Token-count checks are not billed as generation. Unknown final usage stays unbilled for review. An interrupted request can retain a budget reservation.
• The account ledger and CSV export are the billing record. A coding client's own estimated price may not match this service.
• Images, file attachments, hosted search, background jobs, stored-response references, and extended thinking are not supported. Clients must send conversation history. Reading a file locally and sending its text through a tool is supported.

Troubleshooting

Configuration edition: October 23, 2026. Read → edit → verify flows and a six-test code repair task were checked on macOS with Codex 0.163.0-alpha.11.2 and Claude Code 2.1.308. Windows instructions and other cloud/editor clients are provided for setup and are not yet end-to-end verified.