# Providers Bring any model from any provider. zot ships a broad provider catalog so `/login`, `/model`, `--provider`, and `--list-models` all speak the same IDs. Use `/login` for subscription auth and simple API-key providers. Providers that need cloud-specific settings show setup guidance instead of opening a localhost browser form, which is safer for SSH, containers, and `kubectl exec` sessions. ## HTTP proxy Set one global proxy for zot-managed HTTP and HTTPS traffic with `http_proxy` in `$ZOT_HOME/config.json`: ```json title="$ZOT_HOME/config.json" { "http_proxy": "http://127.0.0.1:7890" } ``` zot applies the setting at startup. Existing `HTTP_PROXY`, `HTTPS_PROXY`, `http_proxy`, and `https_proxy` environment variables take precedence for their corresponding protocol. `NO_PROXY` and `no_proxy` continue to control bypasses. Restart zot after changing the setting. > **Proxy credentials** > > `config.json` is not a credential store. If the proxy URL contains a username or password, prefer protected environment variables instead of saving those credentials in the file. ## Subscription login Anthropic Claude Pro/Max (`anthropic`), OpenAI Codex / ChatGPT Plus/Pro (`openai-codex`), Kimi Code (`kimi`), and GitHub Copilot (`github-copilot`). OAuth tokens are stored in `$ZOT_HOME/auth.json` and refreshed when supported. ## GitHub Copilot Use `/login` with `github-copilot`, or supply a GitHub token with Copilot access through `COPILOT_GITHUB_TOKEN` or `GITHUB_COPILOT_TOKEN`. The default model is `claude-sonnet-5`: ```shell zot --provider github-copilot --model claude-sonnet-5 ``` zot selects the wire protocol automatically. Claude uses Anthropic Messages with Copilot Bearer authentication, without Claude Code identity injection. GPT, Grok, and MAI use Responses. Gemini and Kimi use Chat Completions without `reasoning_effort`; their reasoning is managed by the endpoint. Requests include user/agent initiator metadata and image-request headers where appropriate. At startup, zot queries Copilot's authenticated `/models` endpoint in the background. The runtime picker is filtered by model-picker visibility, policy state, and tool-call support. Individual accounts with no eligible picker entries can fall back to explicitly enabled policies; other account types cannot. zot never enables model policies automatically. Account availability stays in memory, separate from the six-hour model cache. Until discovery succeeds, catalog entries remain visible; failed refreshes retain the previous snapshot. Background discovery does not execute command-backed credentials. Restart zot after changing accounts. Filtering helps model selection but does not grant or enforce access; GitHub still authorizes every request. > **Catalog versus account access** > > The website lists public catalog entries, not the models enabled for your account. Retired and utility-only models are excluded from the built-in chat catalog. An ID appearing in an API error's available-model list is not a guarantee that chat requests will succeed. Check your plan, organization policies, and GitHub's current model availability when a request is rejected. Sonnet 4.5 was retired on September 1, 2026. Sonnet 4.6 was also retired except for individual annual Copilot Pro and Pro+ subscribers. It is excluded from the general catalog; eligible subscribers can add it through `$ZOT_HOME/models.json`. MAI-Code-1-Flash uses `mai-code-1-flash-picker` and is scheduled to retire on September 10, 2026. Prefer `mai-code-1.1-flash` for new configurations. GPT-5.4 nano is not a Copilot Chat choice; GitHub offers it through the Codex VS Code extension. > **Claude Fable data retention** > > Anthropic retains Fable prompts and outputs by default for safety classifiers. Eligible enterprises can request a time-limited zero-data-retention exception through GitHub. Enabling the model alone does not grant that exception. Review GitHub's terms before sending sensitive data. For current availability, retirement dates, plan exceptions, and retention terms, see [https://docs.github.com/en/copilot/reference/ai-models/supported-models](https://docs.github.com/en/copilot/reference/ai-models/supported-models). The routing and availability filtering described here are available on runtime `main` from commit `6abb410`; older builds may still use the previous catalog and routing. ## API-key providers Simple API-key providers can be configured through `/login` or environment variables. Examples include Anthropic, OpenAI, OpenAI Responses, DeepSeek, Google Gemini, Kimi/Moonshot, Moonshot CN, Groq, Cerebras, xAI, Together AI, Hugging Face Router, OpenRouter, Mistral, Z.AI, Xiaomi/MiMo, MiniMax, Fireworks, Vercel AI Gateway, and OpenCode. | Provider | Environment | | --- | --- | | Anthropic | `ANTHROPIC_API_KEY` | | OpenAI / OpenAI Responses | `OPENAI_API_KEY` | | Google Gemini | `GEMINI_API_KEY` or `GOOGLE_API_KEY` | | OpenRouter | `OPENROUTER_API_KEY` | | Mistral | `MISTRAL_API_KEY` | | Groq | `GROQ_API_KEY` | | xAI | `XAI_API_KEY` | | Fireworks | `FIREWORKS_API_KEY` | | Vercel AI Gateway | `AI_GATEWAY_API_KEY` | | OpenCode | `OPENCODE_API_KEY` | ## API keys from commands To keep an API key in a password manager instead of storing it directly in `auth.json`, configure an `api_key_command` for the provider: ```json title="auth.json" { "anthropic": { "api_key_command": { "program": "op", "args": ["read", "op://Work/Anthropic/credential"], "timeout_ms": 120000 } } } ``` For providers without a dedicated top-level entry, including providers added through `models.json`, put the same credential object under `additional_api_key_creds` using the provider ID. `program` runs directly without a shell, and every argument must be a separate `args` entry. `timeout_ms` is optional and defaults to 120 seconds. zot runs the command only when that provider is selected. Login-status checks and background model discovery do not execute it. Successful output is cached in memory until zot exits and is never written back to disk. Swarm sub-agents inherit the resolved credential from the supervisor over stdin, so concurrent children do not rerun the command and the credential is not placed in child arguments or environment variables. The command must print one non-empty line to stdout; trailing CR/LF characters are removed and output is limited to 64 KiB. Saving a normal key through `/login` replaces the command, and `/logout` removes it. > **Executable configuration** > > Anyone who can modify `auth.json` can cause zot to run a program under your user account. Keep the file user-writable only and do not copy it from an untrusted source. zot does not interpret `!` prefixes or execute command strings through a shell. ## Cloud setup providers These providers need more than a single pasted key, so `/login` shows setup instructions and points to the provider reference instead of starting a local web login flow. | Provider | Required setup | | --- | --- | | Amazon Bedrock | `AWS_BEARER_TOKEN_BEDROCK` or AWS profile/IAM credentials, plus `AWS_REGION`. | | Google Vertex AI | `GOOGLE_CLOUD_API_KEY` or Google Cloud credentials, plus project/location settings when needed. | | Cloudflare Workers AI | `CLOUDFLARE_API_KEY` and `CLOUDFLARE_ACCOUNT_ID`. | | Cloudflare AI Gateway | `CLOUDFLARE_API_KEY`, `CLOUDFLARE_ACCOUNT_ID`, and `CLOUDFLARE_GATEWAY_ID`. | | Azure OpenAI Responses | `AZURE_OPENAI_API_KEY` and `AZURE_OPENAI_BASE_URL`. | ## Amazon Bedrock Bedrock API keys are bearer tokens. Set `AWS_BEARER_TOKEN_BEDROCK` and `AWS_REGION`, or use the standard AWS credential chain with `AWS_PROFILE`, IAM access keys, ECS task roles, or IRSA. ```shell export AWS_BEARER_TOKEN_BEDROCK=bedrock-api-key-... export AWS_REGION=us-east-1 zot --provider amazon-bedrock --model anthropic.claude-sonnet-4-5-20250929-v1:0 ``` Some Bedrock models require regional inference-profile IDs for on-demand throughput, such as `us.` or `eu.` prefixed model IDs. zot rewrites known families automatically where possible, while explicit profile IDs and ARNs are left unchanged. ## Custom and local providers - Local/compatible: Ollama and generic OpenAI-compatible endpoints via `--base-url`. - LM Studio: register a server URL through `/login` to discover local models automatically. No server is probed before registration. - Managed GGUF models: connect the distinct `llama.cpp` provider to a recent llama.cpp router, then use `/llama` for download, load, unload, and removal. - Named custom providers: define a new provider key with its own `baseUrl` and `api` format in `$ZOT_HOME/models.json`. It logs in through `/login` and reads a derived env var such as `MY_COMPANY_API_KEY`. Add `"discover": true` to list its models live from the OpenAI-compatible `/models` endpoint; see `Models` for details. - Private model IDs, deployment aliases, and provider overrides live in `$ZOT_HOME/models.json`. ## LM Studio > **Build requirement** > > Requires a zot build with LM Studio discovery. If `LM Studio` is not listed in `/login`, use the manual OpenAI-compatible setup on the `Models` page. Open `/login`, choose `api key`, select `LM Studio`, and enter the server URL, for example [http://localhost:1234](http://localhost:1234). Enter an API key if the server requires one, otherwise leave it empty. Zot stores this connection under `lmstudio` in `$ZOT_HOME/auth.json`. The server URL is required. Zot does not scan localhost, and an API key alone does not enable discovery. Registration leaves your current provider and model selected. After registration, open `/model` to refresh and select a discovered model. Zot also refreshes after saving the connection, in the background on startup, and when selecting or resuming an LM Studio model. You can select a model directly after registration: ```shell zot --provider lmstudio --model my-local-model ``` Discovery prefers `/api/v1/models` and falls back to `/v1/models` when the native endpoint returns 404, 405, or 501. It lists models without loading or downloading them. Inference with an unloaded model depends on LM Studio's auto-loading settings. There is no `/llama`-style model-management interface for LM Studio. The discovered list stays in memory, separate from the six-hour cloud model cache and llama.cpp's model list. A failed refresh retains the previous snapshot, and `/model` reports the error while keeping other providers accessible. `/logout lmstudio` removes the connection and stops discovery. Background refresh never executes API-key commands. See `Models` for context limits, defaults, and metadata overrides. The existing coding loop, tools, permissions, and built-in provider pricing do not change. ## llama.cpp router zot has first-class support for llama.cpp router mode. This is separate from Ollama and from a single-model llama.cpp server. The router exposes model management endpoints, while inference uses its OpenAI-compatible `/v1` API. ```shell mkdir -p ~/llama-models llama-server \ --models-dir ~/llama-models \ --no-models-autoload \ --jinja \ --host 127.0.0.1 \ --port 8080 \ -ngl 999 \ -c 32768 ``` Start the router without `--model`, `-m`, or `-hf`. Verify that `GET /health` and `GET /models` work, then run `/login`, choose `api key`, select `llama.cpp`, and enter [http://127.0.0.1:8080](http://127.0.0.1:8080) without `/v1`. The optional API key must match the router's `--api-key` value. ```shell export LLAMA_BASE_URL=http://127.0.0.1:8080 export LLAMA_API_KEY=optional-secret ``` - Open `/llama` to inspect router model state and search Hugging Face GGUF repositories. - Choose a quantization, download it with live byte and percentage progress, and explicitly load or unload it. - Press `d` on a router-downloaded cache model to ask the router to remove it after confirmation. - Open `/model` to refresh and select currently loaded llama.cpp models. Unloaded models are intentionally omitted. > **Cache and removal** > > Hugging Face downloads normally live in the router host's Hugging Face cache, not `--models-dir`. Models placed under `--models-dir` or defined by presets cannot be removed through zot. Some llama.cpp versions retain shared repository artifacts such as `mmproj` files after removing a GGUF, so those leftovers must be cleaned up on the router host if no longer needed. See the `models` page for the full custom-provider schema, credential handling, and the `--insecure` TLS option. The model picker only shows models from providers currently available through env vars, `$ZOT_HOME/auth.json`, Kimi CLI fallback, local Ollama, a registered LM Studio server, or loaded models from a configured llama.cpp router.