Claude Code
Point the Claude Code CLI at Prism with Anthropic environment variables.
Codex
Configure a custom Responses API provider in
~/.codex/config.toml.Cursor
Override the OpenAI base URL in Cursor’s model settings.
OpenCode
Configure a static OpenAI-compatible provider and model catalog.
Hermes
Install Prism’s provider plugin and select a discovered model.
OpenClaw
Install Prism’s provider from ClawHub and use native credential storage.
Shared Prism settings
Use the Anthropic base URL without a
/v1 suffix; Anthropic clients append
/v1/messages themselves.
These are the API model IDs. Gemma requires organization access. OpenCode and
OpenClaw qualify them with their local provider name, such as
prism/deepseek-v4.1-flash, while
sending the bare model ID to Prism. Do not infer availability from marketing
copy. Query the public catalog where the matching recipe supports it; otherwise
use the recipe’s explicit model list.
Credentials
Prism does not offer agent self-signup. If the agent has no key, it asks the workspace owner for an existingPRISM_API_KEY and stores it through the
harness’s native secret mechanism, never repeating it to the human.
Wire protocols
Codex sends stateless OpenAI Responses requests toPOST /v1/responses.
Claude Code sends Anthropic Messages requests to POST /v1/messages. The other
recipes send OpenAI Chat Completions requests to POST /v1/chat/completions.
All use bearer authentication and send the API model ID deepseek-v4.1-flash or
gemma-4-31b in the JSON model field.
See OpenAI Responses,
Anthropic Messages, and
Chat Completions for the complete wire
contracts.