OpenAI — Responses API

Plug in an OpenAI API key and Chamade runs your agent on the Responses + Conversations API — native server-side memory, no infra to host. Bring a model + a system prompt, or reference a published stored prompt.

What it is

An agent provider is the AI runtime that powers your agent. With OpenAI, Chamade calls api.openai.com/v1/responses on each event and delivers the reply. Conversation memory is kept natively by OpenAI's Conversations API (a conv_xxx per conversation), so follow-up turns keep context with no work on your side.

The same transport powers Microsoft Foundry and Azure OpenAI — only the endpoint shape differs. For Azure OpenAI, use the Custom (Responses) preset.

Setup

1. Get an API key from platform.openai.com/api-keys — a project key (sk-proj-…) is recommended for least-privilege scoping.
2. In your dashboard → an agent → pick the OpenAI preset.
3. Endpoint is pre-filled (https://api.openai.com/v1/responses) — leave it. Paste your API key, choose a model (gpt-4o-mini, gpt-5-mini, gpt-5, …).
4. The system prompt is pre-filled with Chamade's XML-action convention (edit freely). Create — a tiny probe validates the key.

Acting & reading

By default the model just replies. Chamade ships the output convention in the system prompt, so the model can act by emitting inline tags Chamade executes:

• <write> — reply in the current DM/call · <say> — speak (TTS) in a call · <file> — attach · <accept/> / <refuse/> / <leave/>.
• Cross-target: <dm platform="discord" to="…" account_id="discord:…">text</dm>. Each event carries an [agent context] line with your account ids, so no lookup is needed.

To let the agent read Chamade state mid-reasoning (transcripts, history, fresh state), Chamade auto-injects its hosted MCP (https://mcp.chamade.io/mcp/) into the request with a scoped bearer — nothing to wire. Rule of thumb: read with MCP, act with the XML tags (never both, or actions fire twice).

Pick an existing agent — stored prompts

Instead of typing a model + system prompt, you can reference a published OpenAI prompt (built in the OpenAI dashboard → Prompts, id pmpt_…, versioned — it carries model + instructions + tools). It's OpenAI's equivalent of a hosted agent: the persona lives in your OpenAI account, Chamade just folds in its action convention.

• In the form, open Stored prompt (OpenAI) → paste the pmpt_… id; optionally pin a version (blank = production).
• Chamade then sends prompt: {id, version} and no model / no instructions (a request-level instruction would override the prompt's persona) — the convention rides in input only.
• OpenAI platform only (not Azure). AgentKit workflows (wf_…) are a different runtime and can't be driven server-side yet.

Memory & context

First turn of a conversation mints a conv_xxx via the Conversations API; follow-up turns reference it, so the model keeps full context server-side (first turn ~2–3s, later turns <1s). On a custom Responses-compatible host that lacks /v1/conversations, Chamade silently falls back to previous_response_id chaining.

Good to know

• The agent needs ≥1 platform account attached in Chamade, or its tools/actions reach nothing.
• Billing is standard OpenAI per-token; the conversation chain expands to full context billed as input tokens (server-side state doesn't change billing).
• Use a project key (sk-proj-…) so you can scope and rotate it independently.
• Azure OpenAI deployments → configure via the Custom (Responses) preset (same transport, different URL + headers).