Models and providers
A provider is where Mework sends model requests: an API service you add with a key, your ChatGPT account, or the Claude Code login on your computer. You add providers and their models in Settings, and pick the model and reasoning effort under the composer.
Add a provider
Providers live in Settings → Providers → Model providers. Open it with the Settings button at the bottom of the sidebar, or ⌘, (Ctrl, on Windows).
A fresh install has two built-in providers: Claude Agent, enabled and in use, and OpenAI Codex, off until you sign in with ChatGPT. If this computer's Claude Code is already signed in, Mework fills Claude Agent with the newest Fable, Opus, Sonnet and Haiku models and selects Opus.
To add your own:
- Click Add provider under the provider list.
- Type a Provider name, choose its Chat protocol, and click Add.
- Paste the key into API Key. It is saved when the field loses focus; emptying the field deletes the stored key.
- Check API address; the line under it shows where chat requests go. For Azure OpenAI, AWS Bedrock and Google Vertex AI, click the gear beside it for the identity fields: requests are refused until those marked
*are filled. - Add models.
Keys go to the system's credential store, never into conversations (where). On macOS, choose Always Allow when asked about Mework Safe Storage.
Chat protocols
Pick the API format the service documents, not its brand.
| Chat protocol | Default address | Notes |
|---|---|---|
| OpenAI Responses | https://api.openai.com/v1 |
|
| OpenAI Chat Completions | https://api.openai.com/v1 |
The default choice. |
| Anthropic Messages | https://api.anthropic.com/v1 |
Also for Anthropic-style endpoints. |
| Google Gemini | https://generativelanguage.googleapis.com/v1beta |
|
| xAI | https://api.x.ai/v1 |
|
| Azure OpenAI | none: enter your resource's address | The model ID is the deployment name. Optional api-version. |
| AWS Bedrock | none: leave blank, Mework derives it from the region | Bedrock API key. Needs AWS region. |
| Google Vertex AI | none: leave blank, Mework derives it from the project and location | Vertex API key (express mode). Needs GCP project and GCP location. |
| OpenAI Compatible | none | Other OpenAI-style servers, such as relays and local model servers. |
DeepSeek, for example, works as OpenAI Chat Completions or OpenAI Responses at https://api.deepseek.com, or as Anthropic Messages at https://api.deepseek.com/anthropic/v1.
The API address must use https://. Addresses on your own network (localhost, names ending in .localhost or .local, loopback and private IPs) may use http:// and need no key, as for Ollama or LM Studio. Any other address needs a key, as do Bedrock and Vertex.
Manage providers
- The switch in the provider's header turns it off; its models then leave the model menu.
- The list order is the order of the model menu. Drag a row, or press ⌥↑ / ⌥↓ (Alt↑ / Alt↓ on Windows). Reordering is off while the list is searched or filtered to enabled or disabled providers.
- Deleting a provider also deletes its stored key. The two built-ins cannot be deleted.
- The gear next to the name opens Provider settings, where you rename the provider, change its API format, or edit its identity fields.
Sign in with ChatGPT
OpenAI Codex runs Codex models on your ChatGPT subscription instead of an API key.
- Select OpenAI Codex and click Sign in with ChatGPT.
- Sign in on the page that opens in your browser. The provider is then enabled.
- Click Fetch models and add the models you want. Signing in adds none.
Sign out disables the provider and deletes the sign-in. The built-in Sol and Luna roles need gpt-6.1-sol and gpt-6-luna added here.
Use your Claude Code login
Claude Agent runs a copy of Claude Code that ships with Mework, so you don't need Claude Code installed. It uses this computer's Claude Code login, from a claude.ai subscription or a Console account.
If the panel does not say Signed in to Claude Code:
- Click Open a terminal to sign in and follow the terminal's prompts.
- Return to Mework and click Re-check. Once signed in, the provider is enabled.
Copy command copies a command that runs the same sign-in in any terminal. If you already use Claude Code, signing in there works too, because both share one login.
Mework turns off Claude Code's own tools, CLAUDE.md, settings files, hooks, plugins, MCP servers, compaction and background tasks, and ignores ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL and CLAUDE_CODE_OAUTH_TOKEN. A Context window above 200,000 gives Claude Code its 1M-token budget.
Add models
The model menu offers exactly the models in a provider's Models list. There is no per-model switch; remove models you do not want.
Fetch models opens Discover models, the provider's catalog: click a row to add or remove that model, or use Add group and Add all results. With a blank address, AWS Bedrock and Google Vertex AI have no catalog to fetch, so Fetch models is not offered. Models already in your list keep the values they have; to pick up a changed limit, edit the model, or remove it and add it again.
To add a model by hand, click + (Add a model manually), type the Model ID exactly as the provider expects it, and click Save.
The minus button on a model or group header removes it at once, without confirmation.
Model properties
The gear on a model opens Model properties. Changes apply when you click Save.
| Property | Default | What it changes |
|---|---|---|
| Model ID | required | The model value sent to the provider, and the name in the model menu. |
| Display name | the ID | The name in the settings list. |
| Group | from the ID | The collapsible group in the settings list. |
| Context window | blank | What the context meter measures against. Auto-compact needs it. |
| Maximum output | blank | Output-token limit per request. When blank, Anthropic Messages and AWS Bedrock send 32,000; other protocols leave it to the provider. |
| Cache lifetime (minutes) | 30 | How long after a request the prompt cache counts as warm; meanwhile Mework asks before a setting change discards it. |
| Vision | from the catalog; off when added by hand | Lets images reach the model. When off, a conversation with images cannot be sent on this model. |
| Mid-conversation tools | on for known models | Lets tool discovery, auto-compact, plan mode and newly enabled tools join a running conversation without losing the prompt cache. Works on Anthropic Messages, OpenAI Responses, Azure OpenAI, OpenAI Codex and Claude Agent. Otherwise tools are fixed at the first request: no auto-compact or Tool discovery, and Plan only if it was on from the start. |
| Mid-conversation system prompts | on for known models | Instructions that start partway through, such as plan mode's guidance, arrive as a system message rather than through box. No effect on Gemini, Bedrock, Vertex or Claude Agent. |
| Prompt caching | on | Anthropic Messages only. Off means no caching. |
| Reasoning field form | Encrypted on OpenAI Responses, OpenAI Codex and Azure OpenAI; Plaintext elsewhere | Matters only on those three, where Mework always asks for encrypted reasoning so thinking replays across tool calls; this setting decides how thinking is shown. Choose Plaintext for endpoints that stream readable reasoning text, such as DeepSeek or many relays, to see it as thinking. |
For a relay or a model Mework does not know, set the two Mid-conversation switches by what the endpoint accepts.
Choose a model
The model button under the composer lists the models of every enabled provider.
Note The model you pick applies to the whole app, not only this conversation: the next request of every conversation uses it.
Reasoning effort, beside it, belongs to the conversation: low, medium, high, extra or max, default medium. Your pick also becomes the starting level for new conversations. A model that does not take the chosen level gets the highest one it takes below it, or else its lowest; on Claude Agent, Claude Code uses high instead.
Agent roles can set their own model and Reasoning effort. A role whose model is not in an enabled provider's list cannot be called.