Models & Providers
Overview
A provider is the service Meowbert calls for AI responses. A model is the model ID that service offers. An agent preset is a user-facing choice in the composer that combines a model with defaults such as reasoning effort.
For example, connect a provider once, then offer a Fast preset and a Deep Think preset using models available through it. Users choose the preset when creating a task; background jobs can have separate model settings.
| You want to change… | Configure… |
|---|---|
| Which service receives model requests | The provider's base URL and credentials |
| Which model and defaults people can select | Agent presets |
| Models used for routing, titles or memory | Background models and embeddings |
| A user's own provider instead of the server's | Bring Your Own Provider |
These settings choose the AI connection, not the machine executing commands. A task running locally on a desktop still calls its model through the server.
Providers
Meowbert calls models through the OpenAI Responses API. A provider is just a base URL and an API key, for example:
| Provider | Base URL |
|---|---|
| OpenAI | https://api.openai.com/v1 |
| An API gateway or proxy | Its OpenAI-compatible base URL, if it supports /v1/responses |
| A local model server | For example http://192.168.1.20:8000/v1, if it supports the Responses API. Use an address the containers can reach, not localhost |
The setup wizard adds your first provider. Later, manage providers in Admin Panel → AI providers. You can save several and switch between them. The selected provider is used for all platform model calls.
Chat Completions-only servers
Providers that only implement /v1/chat/completions don't work yet. Check that your provider or local server supports /v1/responses.
Models people can choose
The composer's model choices are agent presets: a model ID plus defaults such as reasoning effort. Open Admin Panel → Model → Agent presets to configure them. See Agent presets for JSON examples and field definitions.
Agent Swarm presets
A swarm preset adds a leader, optional model router, delegation effort, budgets and lane limits. These defaults also appear in the task composer. The Agent Swarm preset reference explains how the settings interact.
Background models
Admin Panel → Model → Specialized runtime sets the models Meowbert uses behind the scenes:
| Setting | Used for | If unset |
|---|---|---|
internalModel | Routing connector messages to the right project, and as the fallback task model | openai.defaultModel from the config file |
fastModel | Task titles, summarizing long conversations, and other quick jobs inside a task | The task's own model |
reviewerAgent | The reviewer in Quality control tasks (a preset ID) | Built-in default |
memorySynthesisAgent | Memory refresh tasks (a preset ID) | Built-in default |
subagentFastAgent | Fast subagents (a preset ID) | Not available |
A small, fast model is a good choice for internalModel and fastModel.
Embeddings
Two features use an embedding model: Workspace Memory search, and session history search, which lets agents search earlier context windows of a task by meaning. Both are set up in Admin Panel → Platform → Embeddings.
Features
Each feature has its own switch at the top of the tab. Changes save right away and apply from each task's next run.
| Switch | Default | When it is off |
|---|---|---|
| Memory search | On | Memory is not indexed and agents get no memory_search tool. Memory files are still read and written. |
| Session history search | Off | Agents get no history_search tool. They can still read earlier windows from the task's history files. |
Session history search embeds every message and tool output of every running task, up to the first 16,000 characters of each item, so it costs more than Memory search. Its index lives on each worker's local disk, next to the Memory index cache, and is built as tasks run. Each worker removes archived and deleted tasks from its index within about a minute; a restored task is indexed again the next time it runs. A worker that was offline for more than 30 days rebuilds its index from scratch.
Model
- Provider — Follow the selected AI provider, or pin one of your saved providers. It must support the OpenAI-compatible
/v1/embeddingsendpoint. - Model — The embedding model ID. Defaults to
text-embedding-3-small. - Prompt templates — Only for models that expect a prefix, such as
search_query: {text}. Leave them empty otherwise.
The session history index rebuilds by itself when these settings change. The rest of this section is about Memory.
Saving these settings never re-embeds Memory, so you can switch providers that serve the same model without rebuilding anything. After changing the model or templates, the tab shows that the indexes are out of date, and search quality drops until you re-embed. If the new model returns a different vector size, Memory search stops working until you do:
- On next use marks every index for rebuilding. Each workspace re-embeds the next time it's used, and its first Memory search waits for that.
- Wipe & re-index rebuilds every workspace now. Choose how many to rebuild at once, then follow the progress on the same page. You can cancel it, and one interrupted by a worker restart picks up again on its own.
If a provider serves different weights under the same model name, re-embed after switching to it.
Upgrading
Embedding settings are no longer read from config/global.json. If you had a memory.embeddings block there, set the same provider and model in the Embeddings tab, then remove the block.
Context windows and model details
Meowbert needs to know each model's context window to decide when to compact long conversations. Add entries in Admin Panel → Model → Model metadata. See Context Compaction.
Code mode is on for every model by default. If a model handles it poorly, add "code_mode": false to that model's entry, or to the default entry to turn it off for all models. A workspace only uses code mode when both the model and the workspace allow it.
Fallback models
For tasks using context management v2, fallbackModels lets a run continue on another model when its model requests exhaust retries and recovery. Meowbert opens a fresh context window with the recent user messages, unread tool results, saved notes, and access to earlier history. A system note tells the fallback model to continue or complete the task.
Add an ordered list in Admin Panel → Model → Model metadata. These are model IDs available through the same provider, rather than agent preset or router IDs. For example, replace these example IDs with models your provider offers:
{
"primary-model": {
"type": "openai",
"context_management": "v2",
"context_window": 256000,
"fallbackModels": ["backup-model", "second-backup"]
},
"backup-model": {
"type": "claude",
"context_management": "v2",
"context_window": 200000
},
"second-backup": {
"type": "google",
"context_management": "v2",
"context_window": 1000000
}
}The original model's list is tried in order, once per model per run. Each fallback uses its own type, compatibility settings, context window, and usage rates. The active request payload stays the same, including reasoning effort and other settings from the workspace, agent preset, or router target. Targets configured for v1 are skipped. Put fallbackModels on default to provide a shared list; an empty list on a model disables that default for it.
Cancellation, quota limits, and lost or uncertain run operations still stop the run. Fallbacks do not change the task's selected model for future runs.
Letting users bring their own provider
Users can connect their own API key or, experimentally, a ChatGPT subscription in their settings. Their tasks then use their provider instead of yours. See Bring Your Own Provider.