These examples use the website API host, which currently lists their recommended models. The direct API host can have a different catalog during rollouts. For larger requests or longer runtimes, use https://api.nano-gpt.com/api/v1 with a model from that host’s catalog. See API Hosts for limits and availability checks.

OpenCode Integration

Use NanoGPT with OpenCode, the open source AI coding agent for your terminal.

Requirements

OpenCode installed (curl -fsSL https://opencode.ai/install | bash)
A NanoGPT account with API access

One-Click Setup (Mac/Linux)

Run this command in your terminal:

curl -fsSL "https://nano-gpt.com/install/opencode_nanogpt.sh" | bash

This will:

Open your browser to authenticate with NanoGPT (or let you paste an API key)
Configure OpenCode to use NanoGPT as your AI provider
Add starter model definitions while preserving existing model selections, filters, and connection options

OpenCode also includes NanoGPT models from its built-in Models.dev catalogue. You do not need to define each of those models manually. The installer adds starter model definitions; these do not restrict the model list. Python 3 is required to merge existing OpenCode files. The installer stops if an existing JSON file cannot be read, rather than replacing it. It updates opencode.json; if you use opencode.jsonc, follow the manual setup below and merge the settings into your existing file. Setup replaces NanoGPT authentication with your chosen login or pasted key. It removes provider-level options.apiKey and Authorization overrides in provider and model headers from opencode.json so the new credentials in auth.json take effect, while preserving unrelated headers and custom model settings.

Manual Setup

If you prefer to configure manually:

1. Create the auth file

Add your NanoGPT API key to ~/.local/share/opencode/auth.json. If the file already exists, merge the nano-gpt entry into it and keep your other authentication entries:

{
  "nano-gpt": {
    "type": "api",
    "key": "your-nano-gpt-api-key"
  }
}

2. Create the config file

Create ~/.config/opencode/opencode.json, or merge these settings into your existing config:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "nano-gpt/anthropic/claude-sonnet-5.5",
  "provider": {
    "nano-gpt": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "NanoGPT",
      "options": {
        "baseURL": "https://nano-gpt.com/api/v1"
      }
    }
  }
}

This uses OpenCode’s built-in NanoGPT catalogue. A models block is optional for models already in that catalogue. Use it to add a missing model or override model settings, as described in Adding More Models. OpenCode also supports opencode.jsonc for configs with comments. The examples here use strict JSON.

Usage

After setup, start OpenCode:

opencode

Switching Models

Use the /models command inside OpenCode to switch between models, or edit the model field in your opencode.json:

{
  "model": "nano-gpt/deepseek/deepseek-v4-pro"
}

Available Models

Examples of NanoGPT models are listed below. OpenCode’s built-in catalogue and your custom definitions determine which models appear in /models; the list is not limited to entries in your models block.

Model ID	Description
`nano-gpt/anthropic/claude-sonnet-5.5`	Claude Sonnet 5.5 (default)
`nano-gpt/anthropic/claude-opus-5.5`	Claude Opus 5.5
`nano-gpt/openai/gpt-6.1-sol`	GPT 6.1 Sol
`nano-gpt/deepseek/deepseek-v4-pro`	DeepSeek V4 Pro
`nano-gpt/deepseek/deepseek-v4-pro:thinking`	DeepSeek V4 Pro Thinking
`nano-gpt/deepseek/deepseek-v4-flash`	DeepSeek V4 Flash
`nano-gpt/deepseek/deepseek-v4-flash:thinking`	DeepSeek V4 Flash Thinking
`nano-gpt/google/gemini-flash-latest`	Gemini Flash Latest
`nano-gpt/google/gemini-3.1-pro-preview`	Gemini 3.1 Pro Preview
`nano-gpt/z-ai/glm-5.3`	GLM 5.3

Adding More Models

The models block adds model definitions or overrides settings for existing models. It does not hide models that you have not listed. If a model already appears in OpenCode and you do not need custom settings, you can omit its entry. Keep entries for missing models or custom names, options, variants, and context/output limits. To add or override a model, merge its definition into provider.nano-gpt.models:

{
  "provider": {
    "nano-gpt": {
      "models": {
        "z-ai/glm-5.3": {
          "name": "GLM 5.3",
          "limit": { "context": 1048576, "output": 131072 }
        }
      }
    }
  }
}

List NanoGPT API models with GET https://nano-gpt.com/api/v1/models. OpenCode’s catalogue can lag behind the API, so a model missing from OpenCode may need a manual definition. A model appearing in the API list does not necessarily support the tool calls needed by OpenCode.

Limiting the Model List

To show only selected NanoGPT models, add whitelist under provider.nano-gpt. Merge this example into your existing config, keeping your connection settings and any model overrides:

{
  "model": "nano-gpt/anthropic/claude-sonnet-5.5",
  "provider": {
    "nano-gpt": {
      "whitelist": [
        "anthropic/claude-sonnet-5.5",
        "z-ai/glm-5.3"
      ]
    }
  }
}

Use model IDs without the nano-gpt/ prefix inside whitelist, blacklist, and models. The top-level model field uses the full nano-gpt/model-id format. Alternatively, blacklist hides specific models while keeping the rest. If both filters are set, a model must be in whitelist and must not be in blacklist. Keep the selected default model inside your allowed list, then restart OpenCode after saving.

Claude Prompt Caching

OpenCode adds explicit cache_control markers to Claude requests automatically. Its markers omit the TTL and use the five-minute default unless you request a different TTL. You do not need a promptCaching block or an anthropic-beta header just to use OpenCode’s default Claude caching. A cache marker requests caching; it does not guarantee a cache hit. The prompt must meet the selected model’s minimum cacheable size, and later requests must reuse the same prefix before the cache expires.

On NanoGPT’s Chat Completions endpoint, supported explicit cache controls use pay-as-you-go billing, including markers OpenCode adds automatically. They bypass subscription coverage for that request. See subscription billing for prompt caching.

One-Hour Caching

To request one-hour caching for Opus 5.5, merge this configuration into your existing opencode.json. Keep your NanoGPT authentication in auth.json:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "nano-gpt/anthropic/claude-opus-5.5",
  "provider": {
    "nano-gpt": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "NanoGPT",
      "options": {
        "baseURL": "https://nano-gpt.com/api/v1"
      },
      "models": {
        "anthropic/claude-opus-5.5": {
          "options": {
            "promptCaching": {
              "enabled": true,
              "ttl": "1h"
            }
          }
        }
      }
    }
  }
}

OpenCode forwards the model’s options.promptCaching as the top-level promptCaching field in the API request. NanoGPT applies the requested TTL to the cache boundaries, including OpenCode’s existing markers. No additional anthropic-beta header is required. One-hour caching has a higher cache-write price than five-minute caching on standard Claude cache pricing. Choose it when you expect to reuse a large prefix after pauses longer than five minutes. See cache pricing. The example leaves stickyProvider at its default, false, so NanoGPT can fail over when a service is unavailable. Adding "stickyProvider": true inside promptCaching disables failover and can return a 503 instead. It is optional and does not enable caching by itself.

Which Model Settings Should I Keep?

Setting	Purpose	When to keep it
`name`	Display label in OpenCode	When you want a custom label or are defining a missing model.
`reasoning: true`	Tells OpenCode that the model supports reasoning; it is capability metadata, not an API instruction to enable thinking.	When defining a missing model or correcting stale catalogue metadata. Usually unnecessary for a correctly listed model.
`variants`	Adds selectable settings such as reasoning effort	When you want custom choices or a choice missing from your OpenCode version.
`options.promptCaching`	Sends NanoGPT cache controls	When overriding the default TTL, disabling explicit caching, or setting other cache controls.
`options.provider.zdr`	A request field forwarded by OpenCode; it is not NanoGPT-wide ZDR enforcement	Keep intentional privacy settings. To require NanoGPT ZDR routing, use your account or API key’s ZDR only setting.
`promptCaching.stickyProvider`	Prevents failover for explicit caching	Only when preserving the cache matters more than request availability.
`headers.anthropic-beta`	Alternative way to request Claude cache controls	Unnecessary when using the `promptCaching` helper above.

Removing the entire options block also removes your caching and request-level privacy settings. Removing variants only removes your custom effort choices.

Opus 5.5 Reasoning Effort

Opus 5.5 uses adaptive thinking automatically. You do not need reasoning: true or custom variants to enable its thinking. Supported effort values are low, medium, high, xhigh, and max; none cannot disable this model’s thinking. If your OpenCode version does not offer the choices you want, merge these variants into provider.nano-gpt.models["anthropic/claude-opus-5.5"], alongside the existing options:

{
  "reasoning": true,
  "variants": {
    "low": { "reasoningEffort": "low" },
    "medium": { "reasoningEffort": "medium" },
    "high": { "reasoningEffort": "high" },
    "xhigh": { "reasoningEffort": "xhigh" },
    "max": { "reasoningEffort": "max" }
  }
}

OpenCode maps the selected variant’s reasoningEffort to NanoGPT’s reasoning_effort request field. See reasoning controls for the distinction between generating thinking and displaying it. The snippet includes reasoning: true so OpenCode can expose reasoning choices even with missing or stale catalogue metadata. You can omit it when the catalogue already describes the model correctly.

Checking Cache Hits

Look for cache_creation_input_tokens on an eligible write and cache_read_input_tokens > 0 on a subsequent request that reuses the prefix. A successful answer alone does not establish a cache hit or the requested TTL. Raw API responses can also include x_nanogpt_pricing.cacheTTL for the billing TTL. If your OpenCode version does not display these fields, use the API cache verification examples. A short prompt can return zero cached tokens even with a correct configuration. The request format above was checked with OpenCode 1.14.48 and @ai-sdk/openai-compatible. Other SDKs or older client versions can map options differently.

Troubleshooting

OpenCode not finding the config? Use opencode.json or opencode.jsonc, without a leading dot:

Global config: ~/.config/opencode/opencode.json, or $XDG_CONFIG_HOME/opencode/opencode.json when XDG_CONFIG_HOME is set.
Project config: opencode.json in your project.

OpenCode merges config sources. Project settings can override global settings, and a custom config set with OPENCODE_CONFIG can add another source. See OpenCode’s config documentation for the full precedence rules. More models appear than expected? Check that whitelist is inside provider.nano-gpt, not inside models. Defining a models block does not filter the catalogue. Also check project configs and opencode.jsonc for settings that override your global filters. The installer cannot merge your config? Install Python 3 when updating existing OpenCode files. The installer reads strict JSON only; if your config contains comments, use manual setup in opencode.jsonc. The installer stops before authentication or file updates if existing auth or JSON config files cannot be read or parsed. Authentication errors? Check that your API key is correctly set in ~/.local/share/opencode/auth.json (as an object with type and key) and that the key name matches the provider ID (nano-gpt). Claude caching is not being used? If you configured a TTL override, check that promptCaching is inside the selected model’s options, that enabled is true, and that the TTL is exactly "5m" or "1h". Keep reusable instructions and tool definitions at the beginning of the prompt. Changes to that prefix, insufficient prompt length, expiration, or failover can prevent a cache hit. A request returns “temporarily unavailable”? If you set promptCaching.stickyProvider: true, a 503 can mean failover was prevented to preserve the cache. Privacy and provider restrictions can also reduce the available routes. Keep intentional privacy requirements; a service error alone does not establish that those requirements caused the failure. Retry, and send support the new request ID, model ID, and approximate time if the error continues.

Get Started

Endpoint Examples

API Reference

Miscellaneous

Partner

Integrations

OpenCode

OpenCode Integration

Requirements

One-Click Setup (Mac/Linux)

Manual Setup

1. Create the auth file

2. Create the config file

Usage

Switching Models

Available Models

Adding More Models

Limiting the Model List

Claude Prompt Caching

One-Hour Caching

Which Model Settings Should I Keep?

Opus 5.5 Reasoning Effort

Checking Cache Hits

Troubleshooting

​OpenCode Integration

​Requirements

​One-Click Setup (Mac/Linux)

​Manual Setup

​1. Create the auth file

​2. Create the config file

​Usage

​Switching Models

​Available Models

​Adding More Models

​Limiting the Model List

​Claude Prompt Caching

​One-Hour Caching

​Which Model Settings Should I Keep?

​Opus 5.5 Reasoning Effort

​Checking Cache Hits

​Troubleshooting

OpenCode Integration

Requirements

One-Click Setup (Mac/Linux)

Manual Setup

1. Create the auth file

2. Create the config file

Usage

Switching Models

Available Models

Adding More Models

Limiting the Model List

Claude Prompt Caching

One-Hour Caching

Which Model Settings Should I Keep?

Opus 5.5 Reasoning Effort

Checking Cache Hits

Troubleshooting