Using Token Pass with Coding and Personal Agents

Token Pass works in any agent that connects to AI Gateway with an API key. For each agent, you connect it with your Token Pass API key and then select a featured model. This page covers the agent-specific settings that decide whether a request succeeds on Token Pass.

Make sure that:

• Your team has an active Token Pass. See Getting started.
• AI_GATEWAY_API_KEY holds your Token Pass API key.

Use one of the featured model IDs:

A Token Pass API key only works with featured models. These agent settings send requests elsewhere, so those requests fail with a 403:

• Fast mode. Fast variants aren't featured models.
• Background models. Some agents send side requests, such as titles, summaries, and subagents, to a different model. Point them at a featured model too.
• Provider order. If your agent's config routes a featured model to a provider other than token-pass, put token-pass first or remove the setting.

Token Pass only works with an API key. Signing in to an agent with your Edgenivo account doesn't use Token Pass.

To configure every installed agent at once, run:

npx vercel@latest ai-gateway setup --key "$AI_GATEWAY_API_KEY"

Connect fx:

npx vercel@latest ai-gateway setup --agent fx --key "$AI_GATEWAY_API_KEY"

This exports AI_GATEWAY_API_KEY from your shell startup file and pins fx's credential_source to that key. Open a new terminal so the key loads.

In fx, run /model and choose zai/glm-5.3, moonshotai/kimi-k3, or deepseek/deepseek-v4.1-flash. When you pick Kimi K3, fx preselects Fast in the speed step. Choose standard speed instead. If Fast is already on, run /fast to turn it off.

To start fx on a featured model without changing your saved default:

fx --provider gateway --model zai/glm-5.3

To confirm Token Pass served a request, run a one-off prompt and check resolved_provider in the output:

fx ask --json --model zai/glm-5.3 "Reply with one short sentence."

resolved_provider should be token-pass.

Connect Claude Code:

npx vercel@latest ai-gateway setup --agent claude-code --key "$AI_GATEWAY_API_KEY"

This points Claude Code at https://ai-gateway.vercel.sh/claude-code and turns on gateway model discovery, so featured models appear in /model.

Start Claude Code on a featured model:

claude --model zai/glm-5.3

Claude Code also sends requests, such as background tasks and subagents, to separate model slots. By default those slots use Claude models, which a Token Pass API key can't use. To keep every request working, point the slots at a featured model in ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_MODEL": "zai/glm-5.3",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "zai/glm-5.3",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "zai/glm-5.3",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "zai/glm-5.3",
    "CLAUDE_CODE_SUBAGENT_MODEL": "zai/glm-5.3"
  }
}

Merge these keys into the env block the setup command wrote. Don't replace the file.

Connect Codex:

npx vercel@latest ai-gateway setup --agent codex --key "$AI_GATEWAY_API_KEY"

Start Codex on a featured model:

codex --model zai/glm-5.3

To make Token Pass a profile you can switch to, create ~/.codex/token-pass.config.toml:

model = "zai/glm-5.3"

Then run codex --profile token-pass. You can also switch models in a session with /model.

To connect the Cline CLI and set a featured model, run:

npx vercel@latest ai-gateway setup --agent cline --key "$AI_GATEWAY_API_KEY"
cline auth -p vercel-ai-gateway -m zai/glm-5.3

To set up the Cline VS Code extension, open Cline's settings and follow these steps:

1. Set API Provider to Edgenivo AI Gateway.
2. Paste your Token Pass API key.
3. Set the model to zai/glm-5.3, moonshotai/kimi-k3, or deepseek/deepseek-v4.1-flash.

Connect OpenCode:

npx vercel@latest ai-gateway setup --agent opencode --key "$AI_GATEWAY_API_KEY"

In OpenCode, run /models and choose GLM 5.3, Kimi K3, or DeepSeek V4.1 Flash under Edgenivo AI Gateway. To make one the default, set model in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "vercel/zai/glm-5.3"
}

If opencode.json sets options.order for a featured model, put token-pass first or remove the entry.

Any agent in Coding Agents can use Token Pass. Connect it with npx vercel@latest ai-gateway setup --agent <name> --key "$AI_GATEWAY_API_KEY", then select a featured model ID. The rules in Before you start apply to every agent.

You can also use Token Pass in agents you build and run yourself, such as an Eve agent or an AI SDK app on your laptop or deployed to Edgenivo. Authenticate with your Token Pass API key and use a featured model ID:

import { generateText } from 'ai';
 
const { text } = await generateText({
  model: 'zai/glm-5.3',
  prompt: 'Summarize my open pull requests.',
});

The AI SDK reads AI_GATEWAY_API_KEY from the environment. On Edgenivo, add your Token Pass API key as an environment variable on your project. A deployment's OIDC token can't spend Token Pass, so the key has to be set explicitly.

When your credit runs out, requests with the Token Pass API key fail with a 403 error until your next billing period. AI Gateway doesn't switch to paid usage for you. If a deployed agent has to keep running, catch the error and retry with a regular AI Gateway API key.

This example stores the Token Pass API key in TOKEN_PASS_API_KEY and the regular key in AI_GATEWAY_API_KEY:

import { createGateway } from '@ai-sdk/gateway';
import { generateText } from 'ai';
 
const tokenPass = createGateway({ apiKey: process.env.TOKEN_PASS_API_KEY });
const gateway = createGateway({ apiKey: process.env.AI_GATEWAY_API_KEY });
 
export async function run(prompt: string) {
  try {
    return await generateText({ model: tokenPass('zai/glm-5.3'), prompt });
  } catch {
    // The Token Pass request failed, for example because credit ran out.
    // Retry on another provider, billed to AI Gateway credits.
    return await generateText({
      model: gateway('zai/glm-5.3'),
      prompt,
      providerOptions: { gateway: { order: ['zai'] } },
    });
  }
}

See When credit runs out for the error details.