| [Amazon Bedrock](https://aws.amazon.com/bedrock/) | Offers a variety of foundation models, including Claude, Jurassic-2, and others. **AWS environment variables must be set in advance, not configured through `goose configure`** | Credential auth: `AWS_PROFILE`, or `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, `AWS_REGION`<br /><br />Bearer token auth: `AWS_BEARER_TOKEN_BEDROCK` and `AWS_REGION`, `AWS_DEFAULT_REGION`, or `AWS_PROFILE` |
| [Amazon SageMaker TGI](https://docs.aws.amazon.com/sagemaker/latest/dg/realtime-endpoints.html) | Run Text Generation Inference models through Amazon SageMaker endpoints. **AWS credentials must be configured in advance.** | `SAGEMAKER_ENDPOINT_NAME`, `AWS_REGION` (optional), `AWS_PROFILE` (optional) |
| [Anthropic](https://www.anthropic.com/) | Offers Claude, an advanced AI model for natural language tasks. | `ANTHROPIC_API_KEY`, `ANTHROPIC_HOST` (optional) |
| [Avian](https://avian.io/) | Cost-effective inference API with DeepSeek, Kimi, GLM, and MiniMax models. OpenAI-compatible with streaming and function calling support. | `AVIAN_API_KEY`, `AVIAN_HOST` (optional) |
| [ChatGPT Codex](https://chatgpt.com/codex) | Access GPT-5 Codex models optimized for code generation and understanding. **Requires a ChatGPT Plus/Pro subscription.** | No manual key. Uses browser-based OAuth authentication for both CLI and Desktop. |
| [Databricks](https://www.databricks.com/) | Unified data analytics and AI platform for building and deploying models. | `DATABRICKS_HOST`, `DATABRICKS_TOKEN` |
| [Docker Model Runner](https://docs.docker.com/ai/model-runner/) | Local models running in Docker Desktop or Docker CE with OpenAI-compatible API endpoints. **Because this provider runs locally, you must first [download a model](#local-llms).** | `OPENAI_HOST`, `OPENAI_BASE_PATH` |
| [GCP Vertex AI](https://cloud.google.com/vertex-ai) | Google Cloud's Vertex AI platform, supporting Gemini and Claude models. **Credentials must be [configured in advance](https://cloud.google.com/vertex-ai/docs/authentication).** Filters for allowed models by organization policy (if configured). | `GCP_PROJECT_ID`, `GCP_LOCATION` and optionally `GCP_MAX_RATE_LIMIT_RETRIES` (5), `GCP_MAX_OVERLOADED_RETRIES` (5), `GCP_INITIAL_RETRY_INTERVAL_MS` (5000), `GCP_BACKOFF_MULTIPLIER` (2.0), `GCP_MAX_RETRY_INTERVAL_MS` (320_000). |
| [GitHub Copilot](https://docs.github.com/en/copilot/using-github-copilot/ai-models) | Access to AI models from OpenAI, Anthropic, Google, and other providers through GitHub's Copilot infrastructure. **GitHub account with Copilot access required.** | No manual key. Uses [device flow authentication](#github-copilot-authentication) for both CLI and Desktop. |
| [LM Studio](https://lmstudio.ai/) | Run local models with LM Studio's OpenAI-compatible server. **Because this provider runs locally, you must first [download a model](#local-llms).** | None required. Connects to local server at `localhost:1234` by default. |
| [Ollama](https://ollama.com/) | Local model runner supporting Qwen, Llama, DeepSeek, and other open-source models. **Because this provider runs locally, you must first [download and run a model](#local-llms).** | `OLLAMA_HOST` |
| [Ollama Cloud](https://ollama.com/) | Access hosted models on ollama.com via OpenAI-compatible API. Requires an Ollama account and API key. | `OLLAMA_CLOUD_API_KEY` |
| [OpenRouter](https://openrouter.ai/) | API gateway for unified access to various models with features like rate-limiting management. | `OPENROUTER_API_KEY` |
| [OVHcloud AI](https://www.ovhcloud.com/en/public-cloud/ai-endpoints/) | Provides access to open-source models including Qwen, Llama, Mistral, and DeepSeek through AI Endpoints service. | `OVHCLOUD_API_KEY` |
| [Ramalama](https://ramalama.ai/) | Local model using native [OCI](https://opencontainers.org/) container runtimes, [CNCF](https://www.cncf.io/) tools, and supporting models as OCI artifacts. Ramalama API is a compatible alternative to Ollama and can be used with the goose Ollama provider. Supports Qwen, Llama, DeepSeek, and other open-source models. **Because this provider runs locally, you must first [download and run a model](#local-llms).** | `OLLAMA_HOST` |
| [Snowflake](https://docs.snowflake.com/user-guide/snowflake-cortex/aisql#choosing-a-model) | Access the latest models using Snowflake Cortex services, including Claude models. **Requires a Snowflake account and programmatic access token (PAT)**. | `SNOWFLAKE_HOST`, `SNOWFLAKE_TOKEN` |
| [VMware Tanzu Platform](https://techdocs.broadcom.com/us/en/vmware-tanzu/platform/ai-services/10-3/ai/index.html) | Enterprise-managed LLM access through AI Services on VMware Tanzu Platform. Models are fetched dynamically from the endpoint. | `TANZU_AI_API_KEY`, `TANZU_AI_ENDPOINT` |
| [Tetrate Agent Router Service](https://router.tetrate.ai) | Unified API gateway for AI models including Claude, Gemini, GPT, open-weight models, and others. Supports PKCE authentication flow for secure API key generation. | `TETRATE_API_KEY`, `TETRATE_HOST` (optional) |
| [Venice AI](https://venice.ai/home) | Provides access to open source models like Llama, Mistral, and Qwen while prioritizing user privacy. **Requires an account and an [API key](https://docs.venice.ai/overview/guides/generating-api-key)**. | `VENICE_API_KEY`, `VENICE_HOST` (optional), `VENICE_BASE_PATH` (optional), `VENICE_MODELS_PATH` (optional) |
goose automatically enables Anthropic's [prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching) when using Claude models via Anthropic, Amazon Bedrock, Databricks, OpenRouter, and LiteLLM providers. This adds `cache_control` markers to requests, which can reduce costs for longer conversations by caching frequently-used context. See the [provider implementations](https://github.com/aaif-goose/goose/tree/main/crates/goose/src/providers) for technical details.
| [Cursor Agent](https://docs.cursor.com/en/cli/overview) (`cursor-agent`) | Uses Cursor's AI CLI tool with your Cursor subscription. Provides access to GPT-5, Claude 4, and other models through the cursor-agent command-line interface. | cursor-agent CLI installed and authenticated |
goose supports [Agent Client Protocol (ACP)](https://agentclientprotocol.com/) agents as providers. ACP providers pass goose extensions through to the agent as MCP servers.
| [Claude ACP](https://github.com/zed-industries/claude-agent-acp) (`claude-acp`) | Uses Claude Code via ACP. Passes goose extensions to the agent as MCP servers. | `npm install -g @zed-industries/claude-agent-acp`, active Claude Code subscription |
| [Codex ACP](https://github.com/zed-industries/codex-acp) (`codex-acp`) | Uses OpenAI Codex via ACP. Passes goose extensions to the agent as MCP servers. | `npm install -g @zed-industries/codex-acp`, active ChatGPT Plus/Pro subscription |
To configure your chosen provider, see available options, or select a model, visit the `Models` tab in goose Desktop or run `goose configure` in the CLI.
2. goose will open a browser window for you to sign in with the credentials of your active ChatGPT Plus or Pro subscription.
3. Authorize goose to access your ChatGPT subscription.
4. When you return to goose Desktop, you're ready to begin your first session.
</TabItem>
<TabItem value="tetrate" label="Agent Router">
We recommend new users start with Agent Router by Tetrate. Tetrate provides access to multiple AI models with built-in rate limiting and automatic failover.
You'll receive $10 in free credits the first time you automatically authenticate with Tetrate through goose. This offer is available to both new and existing Tetrate users.
2. Find the provider of your choice and click its `Configure` button. If you don't see your provider in the list, click `Add Custom Provider` at the bottom of the window to [configure a custom provider](#configure-custom-provider).
3. Depending on your provider, you'll need to input your API Key, API Host, or other optional [parameters](#available-providers). Click the `Submit` button to authenticate and begin your first session.
`goose configure` doesn't support entering custom model names. To use a model not in the provider's list, use goose Desktop or edit the `GOOSE_MODEL` variable in your [`config.yaml`](/docs/guides/config-files) directly.
Need to connect to multiple OpenAI-compatible endpoints? [Configure custom providers](#configure-custom-provider) instead for easier switching and better organization, as well as custom naming and shareable configurations.
| `OPENAI_CUSTOM_HEADERS` | No | Additional headers to include in the request. Can be set via environment variable, configuration file, or CLI, in the format `HEADER_A=VALUE_A,HEADER_B=VALUE_B`. |
For enterprise deployments, you can pre-configure these values using environment variables or configuration files to ensure consistent governance across your organization.
Create custom providers to connect to services that aren't [already supported](#available-providers) or customize how you connect to them. Custom providers appear in goose's provider list and can be selected like any other provider.
Custom providers must use OpenAI, Anthropic, or Ollama compatible API formats. They can include custom headers for additional authentication, API keys, tokens, or tenant identifiers. Each custom provider maps to a JSON configuration file.
- **API Key**: The API key, which is accessed using a custom environment variable and stored in the keychain (or `secrets.yaml` if the keyring is disabled or cannot be accessed)
- For providers that don't require authorization (e.g., local models like Ollama, vLLM, or internal APIs), uncheck the **"This provider requires an API key"** checkbox
- If Yes: You'll be prompted to enter your **API Key** (stored securely in the keychain, or in `secrets.yaml` if the keyring is disabled or cannot be accessed)
- **Custom Headers**: Any additional header names and values
:::info Custom Headers
Currently, custom headers can only be defined for OpenAI compatible providers in the CLI. For Anthropic or Ollama compatible providers, edit the provider configuration file after creation.
Then use the `api_key_env` to set the key for your session. For example:
```bash
export CUSTOM_CORP_API_API_KEY="your-api-key"
goose session start --provider custom_corp_api
```
:::tip Keychain Key Storage
If you want to store the API key in the `goose` keychain, update the provider in goose Desktop and enter the key. This provides secure, persistent storage and allows goose to connect natively to the provider.
3. Select the custom provider you want to update and press `Enter`. Use the arrow keys (↑/↓) to move through the options, or start typing to filter the list.
These free options are a great way to get started with goose and explore its capabilities. However, you may need to upgrade your LLM for better performance.
Groq provides free access to open source (open weight) models with high-speed inference. To use Groq with goose, you need an API key from [Groq Console](https://console.groq.com/keys).
For the complete list of supported Groq models, see [groq.json](https://github.com/aaif-goose/goose/blob/main/crates/goose/src/providers/declarative/groq.json).
Google Gemini provides a free tier. To start using the Gemini API with goose, you need an API Key from [Google AI studio](https://aistudio.google.com/app/apikey).
goose is a local AI agent, and by using a local LLM, you keep your data private, maintain full control over your environment, and can work entirely offline without relying on cloud access. However, please note that local LLMs require a bit more set up before you can use one of them with goose.
goose extensively uses tool calling, so models without it can only do chat completion. If using models without tool calling, all goose [extensions must be disabled](/docs/getting-started/using-extensions#enablingdisabling-extensions).
2. In a terminal, run any Ollama [model supporting tool-calling](https://ollama.com/search?c=tools) or [GGUF format HuggingFace Model](https://huggingface.co/search/full-text?q=%22tools+support%22+%2B+%22gguf%22&type=model):
For the Ollama provider, if you don't provide a host, we set it to `localhost:11434`. When constructing the URL, we prepend `http://` if the scheme is not `http` or `https`. Since Ramalama's default port to serve on is 8080, we set `OLLAMA_HOST=http://0.0.0.0:8080`
If you notice that goose is having trouble using extensions or is ignoring [.goosehints](/docs/guides/context-engineering/using-goosehints), it is likely that the model's default context length of 2048 tokens is too low. Use `ramalama serve` to set the `--ctx-size, -c` option to a [higher value](https://github.com/containers/ramalama/blob/main/docs/ramalama-serve.1.md#--ctx-size--c).
The native `DeepSeek-r1` model doesn't support tool calling, however, we have a [custom model](https://ollama.com/michaelneale/deepseek-r1-goose) you can use with goose.
If you notice that goose is having trouble using extensions or is ignoring [.goosehints](/docs/guides/context-engineering/using-goosehints), it is likely that the model's default context length of 4096 tokens is too low. Set the `OLLAMA_CONTEXT_LENGTH` environment variable to a [higher value](https://github.com/ollama/ollama/blob/main/docs/faq.mdx#how-can-i-specify-the-context-window-size).
2. [Enable Docker Model Runner](https://docs.docker.com/ai/model-runner/#enable-dmr-in-docker-desktop)
3. [Pull a model](https://docs.docker.com/ai/model-runner/#pull-a-model), for example, from Docker Hub [AI namespace](https://hub.docker.com/u/ai), [Unsloth](https://hub.docker.com/u/unsloth), or [from HuggingFace](https://www.docker.com/blog/docker-model-runner-on-hugging-face/)
Beyond single-model setups, goose supports [multi-model configurations](/docs/guides/multi-model/) that can use different models and providers for specialized tasks:
- **Lead/Worker Model** - Automatic switching between a lead model for initial turns and a worker model for execution tasks
- **Planning Mode** - Manual planning phase using a dedicated model to create detailed project breakdowns before execution
Some models expose their internal reasoning or "chain of thought" as part of their response. goose automatically captures this reasoning output and makes it available to you. The following models and providers support reasoning output:
| Provider / Model | How It Works |
|---|---|
| **DeepSeek-R1** (via OpenAI, Ollama, OpenRouter, OVHcloud, etc.) | Reasoning captured from the `reasoning_content` field in the API response |
| **Kimi** (via Groq or other OpenAI-compatible endpoints) | Reasoning captured from the `reasoning_content` field in the API response |
| **Gemini CLI** (Google Gemini models with thinking enabled) | Thinking blocks captured from the streaming response |
| **Claude** (Anthropic, with [Claude thinking](/docs/guides/environment-variables#claude-thinking-configuration) enabled) | Thinking blocks captured from the API response |
Reasoning output appears automatically in a collapsible **"Show reasoning"** toggle above the model's response. Click it to expand and view the model's thought process.
</TabItem>
<TabItem value="cli" label="goose CLI">
Reasoning output is **hidden by default** in the CLI. To display it, set the `GOOSE_CLI_SHOW_THINKING` environment variable:
```bash
export GOOSE_CLI_SHOW_THINKING=1
```
When enabled, reasoning appears under a "Thinking:" header in dimmed text before the model's main response.
:::note
This requires stdout to be a terminal (reasoning output won't appear when piping output to a file or another command).
:::
</TabItem>
</Tabs>
:::tip
Reasoning output can be useful for understanding how the model arrived at its answer, debugging unexpected behavior, or learning from the model's problem-solving approach. However, it can also be verbose — toggle it on only when you need it.
If you have any questions or need help with a specific provider, feel free to reach out to us on [Discord](https://discord.gg/goose-oss) or on the [goose repo](https://github.com/aaif-goose/goose).
