Skip to main content

Tool Setup

Connect each tool with the gateway path it expects, then send one small test request before real work.

Start with these three values​

Every setup starts with:

  • gateway URL
  • Prysm API key
  • exact model name exposed by your gateway

Use the exact model name shown in model marketplace or Model Routes. Tool UIs often show a display label and a request model name separately.

Choose the right setup path​

The gateway supports multiple protocol shapes. The tool determines which path you should use.

ToolBase URL / Env ValueAuthProtocolBest for
Cursor{{BASE_URL}}/cursorCursor OpenAI API Key fieldCursor passthroughAsk, Plan
Claude CodeANTHROPIC_BASE_URL={{BASE_URL}}ANTHROPIC_AUTH_TOKENAnthropic /v1/messages via gateway translationTerminal coding workflows
OpenAI Codex CLIOPENAI_BASE_URL={{BASE_URL}}OPENAI_API_KEYOpenAI /responses via gateway translationCodex CLI agent workflows
OpenCode{{BASE_URL}}/v1OpenCode provider keyOpenAI-compatible /v1Multi-model terminal workflows
Cline{{BASE_URL}}OpenAI Compatible key fieldOpenAI-compatibleVS Code Chat, Plan, Agent
VS Code / Continue{{BASE_URL}}OpenAI-compatible API keyOpenAI-compatibleVS Code assistant workflows
GitHub CopilotVS Code proxy override URL = {{BASE_URL}}editor session authOpenAI-compatible proxy overrideInline completions, chat
Gemini CLIGOOGLE_GEMINI_BASE_URL={{BASE_URL}}GEMINI_API_KEYGemini CLI via gateway routingGemini-style terminal workflows
Qwen Code CLIOPENAI_BASE_URL={{BASE_URL}}OPENAI_API_KEYOpenAI-compatibleQwen-style terminal workflows
Open WebUIconnection URL = {{BASE_URL}}Prysm API keyOpenAI-compatible backendTeam chat workspace
Retool Assistcustom provider base URL = public {{BASE_URL}}Prysm API keyOpenAI-compatible custom providerAI-assisted internal app building
OpenClawprovider base URL = {{BASE_URL}}Prysm API keyOpenAI-compatible provider backendChat app assistants and agents
SDK / Curl{{BASE_URL}} or full endpointBearer keyOpenAI-compatibleApp integration, validation

Common validation first​

Run one request before opening any coding tool. This confirms that the key, model, and gateway path are all valid.

Python SDK​

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_PROXY_KEY",
    base_url="{{BASE_URL}}",
)

response = client.chat.completions.create(
    model="MODEL_NAME",
    messages=[
        {"role": "user", "content": "Reply with the word connected."}
    ],
)

print(response.choices[0].message.content)

JavaScript SDK​

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: "YOUR_PROXY_KEY",
  baseURL: "{{BASE_URL}}",
});

const response = await client.chat.completions.create({
  model: "MODEL_NAME",
  messages: [{ role: "user", content: "Reply with the word connected." }],
});

console.log(response.choices[0].message.content);

Curl​

curl {{CHAT_COMPLETIONS_URL}} \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_PROXY_KEY" \
  -d '{
    "model": "MODEL_NAME",
    "messages": [
      {"role": "user", "content": "Reply with the word connected."}
    ]
  }'

Cursor​

Cursor uses the gateway's /cursor path. The current verified integration supports Ask and Plan modes with Prysm API keys.

  1. Open Cursor → Settings → Cursor Settings → Models.
  2. Fill in OpenAI API Key with your Prysm API key.
  3. Enable Override OpenAI Base URL and set it to {{BASE_URL}}/cursor.
  4. Click Add Custom Model.
  5. Paste the Public Model Name from Model Routes.
  6. Open Ask or Plan and select that model.

First test​

Send a short message in Ask mode such as Reply with: Cursor connected.

Known behavior​

  • Ask and Plan are the primary path for custom key setups.
  • Agent mode follows Cursor's own custom key support boundaries.
  • A Model not responding error usually means the URL is missing /cursor or the key lacks model access.

Claude Code​

Claude Code sends Anthropic Messages API requests. The gateway translates that format to the target provider.

Environment variables​

export ANTHROPIC_BASE_URL="{{BASE_URL}}"
export ANTHROPIC_AUTH_TOKEN="YOUR_PROXY_KEY"

export ANTHROPIC_MODEL="MODEL_NAME"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="MODEL_NAME"
export ANTHROPIC_DEFAULT_OPUS_MODEL="MODEL_NAME"
export ANTHROPIC_DEFAULT_SONNET_MODEL="MODEL_NAME"
export ANTHROPIC_REASONING_MODEL="MODEL_NAME"

Optional config file path​

  • macOS / Linux: ~/.claude/settings.json
  • Windows: C:\Users\your-user-name\.claude\settings.json
{
  "env": {
    "ANTHROPIC_BASE_URL": "{{BASE_URL}}",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_PROXY_KEY",
    "ANTHROPIC_MODEL": "MODEL_NAME",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "MODEL_NAME",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "MODEL_NAME",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "MODEL_NAME",
    "ANTHROPIC_REASONING_MODEL": "MODEL_NAME"
  }
}

First test​

Validate the protocol before opening Claude Code:

curl -X POST {{BASE_URL}}/v1/messages \
  -H "Authorization: Bearer YOUR_PROXY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MODEL_NAME",
    "max_tokens": 128,
    "messages": [
      {"role": "user", "content": "Reply with the word connected."}
    ]
  }'

Start Claude Code with the same model:

claude --model MODEL_NAME

Responses API note​

Recent gateway builds route /v1/messages through /responses by default. This is the recommended forward path.

If your Claude Code request reaches Prysm and still fails at the model layer, ask your Prysm administrator to check the gateway's Anthropic compatibility mode.

gateway_settings:
  use_chat_completions_url_for_anthropic_messages: true

This setting is useful for provider combinations that need chat-completions compatibility behind the gateway.

OpenAI Codex CLI​

OpenAI Codex CLI uses the gateway root URL with standard OpenAI-compatible auth variables.

Version note​

Use a gateway build that already supports /responses routing for Codex workflows.

Environment variables​

export OPENAI_BASE_URL="{{BASE_URL}}"
export OPENAI_API_KEY="YOUR_PROXY_KEY"

Gateway readiness​

Codex works best when Prysm already has:

  • /responses support enabled
  • a stable public model name for Codex to select
  • a gateway configuration that already matches Codex's parameter shape

Ask your Prysm administrator to confirm this if Codex connects to Prysm and still fails before generation starts.

First test​

codex --model gpt-5

Responses API note​

Codex is the clearest place to think in /responses terms. The gateway's /responses support covers streaming, logging, cost tracking, get, delete, and provider bridging across supported models.

OpenCode​

OpenCode works best with the OpenAI-compatible /v1 base.

Configuration file​

Global config usually lives at:

~/.config/opencode/opencode.json

Sample config:

{
  "provider": {
    "prysm": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Prysm",
      "options": {
        "baseURL": "{{BASE_URL}}/v1"
      },
      "models": {
        "gpt-5": {
          "name": "GPT-5"
        },
        "claude-sonnet-4-6-stable": {
          "name": "Claude Sonnet 4.6"
        }
      }
    }
  }
}

Connect and test​

opencode

Inside OpenCode:

/connect

Then:

  • provider name: Prysm
  • API key: your Prysm API key

After that:

/models

Select one configured model and send a short test prompt.

Cline​

Cline works through OpenAI Compatible mode and uses the gateway root path.

  1. Open the project folder in VS Code.
  2. Open Cline in VS Code.
  3. Open settings in the Cline panel.
  4. Select OpenAI Compatible as the provider.
  5. Fill in:
    • Base URL: {{BASE_URL}}
    • API Key: YOUR_PROXY_KEY
    • Model ID: MODEL_NAME
  6. Save and start a chat or plan flow.

First test​

Use a short prompt such as Reply with: Cline connected.

Verified path​

Our internal validation has shown the smoothest Cline experience on OpenAI-compatible models and gateway-exposed domestic models that already pass the base curl test.

VS Code / Continue​

Continue works cleanly when the model config uses an OpenAI-compatible provider for gateway-backed models. This keeps the requested model name aligned with the model identity returned by the gateway.

Add or update a model entry with:

  • provider: OpenAI-compatible
  • base URL: {{BASE_URL}}
  • API key: YOUR_PROXY_KEY
  • model: MODEL_NAME

First test​

Open the same VS Code workspace folder you want Continue to edit, then send a short prompt such as Reply with: Continue connected.

Setup notes​

  • Use OpenAI-compatible config for non-Claude gateway models.
  • Keep Anthropic-specific config for real Claude-family model routes.
  • Open a VS Code workspace folder before asking Continue, Cline, or another extension to create or edit files.

More supported tools​

The gateway also works with several other tools that use OpenAI-compatible, Gemini-style, or editor proxy override flows.

GitHub Copilot​

GitHub Copilot can route through the gateway by overriding the proxy URL in the editor settings.

Add this to VS Code settings.json:

{
  "github.copilot.advanced": {
    "debug.overrideProxyUrl": "{{BASE_URL}}",
    "debug.testOverrideProxyUrl": "{{BASE_URL}}"
  }
}

First test​

Restart VS Code, open Copilot Chat, and send a short message such as Reply with: Copilot connected.

Gemini CLI​

Gemini CLI works through environment variables that point it to the gateway.

export GOOGLE_GEMINI_BASE_URL="{{BASE_URL}}"
export GEMINI_API_KEY="YOUR_PROXY_KEY"

Start the CLI:

gemini

Model routing note​

Gemini CLI is the cleanest fit for Gemini-family model names. If you want Gemini CLI to use another provider behind Prysm, ask your Prysm administrator to expose a compatible model alias for the CLI to call.

First test​

Run a short prompt such as Reply with: Gemini connected.

Qwen Code CLI​

Qwen Code CLI uses the standard OpenAI-compatible environment variables.

export OPENAI_BASE_URL="{{BASE_URL}}"
export OPENAI_API_KEY="YOUR_PROXY_KEY"
export OPENAI_MODEL="qwen-code"

Start the CLI:

qwen

Model routing note​

Qwen Code CLI works best when Prysm exposes a stable public model name such as qwen-code or another exact alias for the CLI to request. Ask your Prysm administrator to confirm the exact model name if needed.

First test​

Run a short prompt such as Reply with: Qwen connected.

Open WebUI​

Open WebUI connects through a standard backend connection.

In Settings → Connections, create a new connection with:

  • URL: {{BASE_URL}}
  • Key: YOUR_PROXY_KEY

First test​

Open the model selector, choose a gateway-exposed model, and send a short message such as Reply with: Open WebUI connected.

Good fit​

Open WebUI is a strong fit for shared chat access, model comparison, and light internal rollout before a team adopts editor-native tools.

Retool Assist​

Retool Assist connects through a custom AI provider.

Hosted Retool requires a public URL. A local localhost gateway URL only works for self-hosted or locally bridged Retool setups.

If you only have an internal Prysm address, ask your Prysm administrator for the public URL that Retool can reach.

In Retool, create a custom AI provider with:

  • Provider Schema: OpenAI
  • Base URL: public {{BASE_URL}}
  • API Key: YOUR_PROXY_KEY
  • Model: exact public model name exposed by the gateway

First test​

Open Retool Assist or an AI resource query and send a short prompt such as Reply with: Retool connected.

Good fit​

Retool Assist is a strong fit for teams that build internal tools and want one governed AI backend for app generation, editing, and query assistance.

OpenClaw​

OpenClaw connects chat channels and agents to the gateway through its onboarding flow.

Run onboarding:

openclaw onboard --install-daemon

During setup:

  • choose the gateway-backed provider option
  • set the base URL to {{BASE_URL}}
  • enter YOUR_PROXY_KEY
  • enter the exact model name exposed by the gateway

First test​

After onboarding, verify with:

openclaw health
openclaw agent --agent main -m "Reply with: OpenClaw connected"

Good fit​

OpenClaw is a strong fit when you want one gateway-backed model layer behind WhatsApp, Telegram, Discord, and other chat surfaces.

SDK and endpoint reference​

Use these values when a tool asks for a full endpoint instead of a base URL:

  • base URL: {{BASE_URL}}
  • chat completions URL: {{CHAT_COMPLETIONS_URL}}
  • embeddings URL: {{EMBEDDINGS_URL}}
  • Anthropic messages URL: {{BASE_URL}}/v1/messages
  • Cursor base URL: {{BASE_URL}}/cursor
  • VS Code extension OpenAI-compatible base URL: {{BASE_URL}}
  • OpenCode /v1 base URL: {{BASE_URL}}/v1

Troubleshooting checklist​

  1. Confirm the tool is using the correct protocol path for that tool.
  2. Confirm the key has access to the selected model.
  3. Confirm the model name matches the gateway-exposed name exactly.
  4. Send one direct test request with curl or SDK before debugging the tool UI.
  5. Open the target VS Code workspace folder before file edits.
  6. Use OpenAI-compatible mode in VS Code extensions for gateway-backed non-Claude models.
  7. Check Logs with the request time to separate key, model, route, and upstream provider errors.