Agentic tools

Agentic coding tools — command-line agents and IDE assistants — run their turns against a model backend. This section shows how to point that backend at DIAL, so a DIAL-hosted model answers through the tool while DIAL applies its roles, quotas, and cost tracking. It is for developers who already have a DIAL host, a DIAL API key, and the tool installed.

Each tutorial is self-contained and ends the same way: a model replying to you through DIAL, so you can see the connection working. Follow the one for the tool you use.

Available tutorials

• Connect Codex to DIAL — the command-line agent that speaks the OpenAI Responses API, direct or through a translating proxy
• Connect GitHub Copilot to DIAL — add your DIAL deployments to the Copilot model picker in VS Code
• Connect Open Claw to DIAL — register DIAL as a model provider in the self-hosted agent gateway
• Connect OpenCode to DIAL — add DIAL as a custom provider in the terminal coding agent
• Connect Claude Code to DIAL — point Anthropic's command-line agent at DIAL's native Anthropic Messages API
• Connect Claude Desktop to DIAL — route the Claude desktop app's Chat, Cowork, and Code sessions through DIAL with third-party inference
• Connect the Claude M365 plugin to DIAL — front DIAL with an NGINX reverse proxy so the Claude for Microsoft 365 Office add-ins can use it as a gateway

Before you begin

Every tutorial needs two things:

• The URL of your DIAL host, written throughout as <YOUR_DIAL_HOST>.
• A DIAL API key, written throughout as <DIAL_API_KEY>.

Each tutorial lists any extra tools it needs at the top. Replace <YOUR_DIAL_HOST> with your real host and <DIAL_API_KEY> with your real key everywhere they appear.

Check your DIAL version

Some behavior depends on your DIAL release. The native Responses API — which lets Codex reach DIAL without a proxy — is only available from DIAL v0.45 onwards, and how deployments handle request settings can vary by release. Ask the host for its version (no key needed):

curl -s https://<YOUR_DIAL_HOST>/version

On Windows, use curl.exe, or Invoke-RestMethod if curl is unavailable:

curl.exe -s https://<YOUR_DIAL_HOST>/version
Invoke-RestMethod https://<YOUR_DIAL_HOST>/version

It prints a bare version string such as 0.45. Below v0.45, no deployment serves the Responses API, so Codex must use the proxy path regardless of any responses_api flag.

Azure-backed deployment settings

Azure OpenAI–backed deployments (for example, GPT-5.x) need two extra request settings. If a call fails with one of these errors, apply the matching fix wherever the tutorial sets the DIAL URL or the output-token limit:

• api-version is a required query parameter — append ?api-version=2025-04-01-preview to the DIAL URL the tool calls (the chat-completions or /responses URL).
• Unsupported parameter: 'max_tokens' … use 'max_completion_tokens' — Azure-backed GPT-5.x reject the plain max_tokens field for any value. Switch the tool to its max_completion_tokens equivalent, or send no output cap at all, since uncapped requests are accepted.

An output cap is never required. You only think about it when a tool sends one: GPT-5.x care about the field name above, while other deployments accept max_tokens but reject a value above that deployment's completion-token ceiling (max_tokens is too large). Not every deployment needs an api-version either, so leave both settings out unless you hit the matching error.

Note

Do not trust a deployment's max_tokens_supported flag here — GPT-5.x report it true yet still reject max_tokens.