Coding agents

Run Prism models in Claude Code and OpenAI-compatible agent harnesses.

Prism exposes OpenAI Chat Completions and Anthropic Messages. Choose the format your coding agent already uses.

Claude Code

Claude Code uses the Anthropic Messages API. Point it at Prism with environment variables:

export ANTHROPIC_BASE_URL="https://api.prisminference.com"
export ANTHROPIC_AUTH_TOKEN="$PRISM_API_KEY"
export ANTHROPIC_MODEL="prism-glm53"
export ANTHROPIC_SMALL_FAST_MODEL="prism-qwen"

claude

To persist the configuration:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.prisminference.com",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_PRISM_API_KEY",
    "ANTHROPIC_MODEL": "prism-glm53",
    "ANTHROPIC_SMALL_FAST_MODEL": "prism-qwen"
  }
}

Save this as ~/.claude/settings.json. Prefer an environment variable or secret injection mechanism over writing a real key directly into the file.

Claude Code tools use the normal Anthropic tool_use and tool_result loop. Streaming, system prompts, and conversation history use the standard Messages format.

OpenAI-compatible harnesses

For an agent that accepts an OpenAI-compatible provider, configure:

base URL: https://api.prisminference.com/v1
API key:  $PRISM_API_KEY
model:    prism-glm53

Use the Chat Completions wire format. Prism V0 does not expose the OpenAI Responses endpoint.

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.PRISM_API_KEY,
  baseURL: "https://api.prisminference.com/v1",
});

const response = await client.chat.completions.create({
  model: "prism-glm53",
  messages,
  tools,
  prompt_cache_key: conversationId,
});

Codex compatibility

Current Codex clients use the OpenAI Responses API. Prism V0 serves Chat Completions and Messages, not Responses, so Codex does not connect directly. Use a gateway that translates Responses requests to Chat Completions, or use an agent harness with native Chat Completions support.

Choose models by role

Agent role	Starting model
Main coding and reasoning loop	prism-glm53
Reasoning-heavy or multimodal work	prism-kimik3
Fast background tasks and summaries	prism-qwen
Fast coding and tool loops	prism-dsv4flash

Measure your own workload before routing production traffic. Context size, tool-schema size, cache reuse, and output length all affect end-to-end latency.