Allowed Models
Amp Enterprise workspace admins can choose which models the members of their workspace can use. Open the workspace settings page, go to “Model Routing”, and then open “Allowed Models”.
Model Groups
Some switches on the “Default Policy” tab turn off a whole group of models, including models a built-in mode uses and models Amp adds to the group later:
- Open-Weights Models turns off every model with published weights, such as GLM, Kimi, DeepSeek, Qwen, and Inkling. Low mode then runs on GPT-5.6 Terra.
- Chinese Labs, inside the Open-Weights Models section, turns off only the models trained by Z.ai, DeepSeek, Moonshot AI, Alibaba, and MiniMax. It is off whenever Open-Weights Models is off.
- SpaceXAI and Meta, on their sections, turn off every Grok or Muse model.
- Claude Fable, inside the Anthropic section, turns off every Claude Fable model. Anthropic retains the prompts and outputs sent to Claude Fable for 30 days, so leaving it on accepts that retention for your workspace.
The group switches save with the rest of the page when you press Save. Directory group grants cannot give back a model whose group is off.
Default Policy
Most models start enabled. On the “Default Policy” tab, turn off the models you do not want members to use and press Save.
Models for authorized security work start disabled. Turn one on for everyone on the “Default Policy” tab, or grant it to a group on the “Group Access” tab.
Members cannot pick a disabled model in the model selector or in the CLI, and Amp refuses to run a thread on it. A thread that was already using a model when you disabled it stops at its next message, and the CLI shows “Model Not Available” with a link to the list of models.
Locked Models
Some models cannot be disabled because Amp currently uses them, so a change on this page can never break new threads. Amp locks the model your workspace’s new threads start on and the models Puck and the Spaces agent run on. The page lists these models under “Used by Amp” with what uses them.
To disable a locked model, open Model Selector in workspace settings. For the model new threads start on, choose a different model under New Threads Start On. With Staff Pick, Last Choice, or Auto, new threads start on the Staff Pick, or on the first Top Model when the workspace’s Top Models leave the Staff Pick out, so choose a fixed model. For Puck or Spaces, choose a different model in its row. Then return to Allowed Models.
A model you disabled before it became the one new threads start on stays disabled. Members then cannot start a thread on it until you choose another model under New Threads Start On.
You cannot pick a disabled model, or a model in a group that is off, for Puck or Spaces. The model menu shows such models greyed out and marked “Disabled”.
Group Access
If your workspace has directory sync enabled (via WorkOS SCIM), the “Group Access” tab lists the groups from your identity provider. You can give the members of a group models that are disabled for everyone else. Press Choose Models on a group, select the models, and save.
Grants only add access. A member of several groups gets every model that any of their groups grants. When you enable a model for everyone, its grants are removed. When a group is deleted in your identity provider, its grant is listed under “Groups No Longer in Your Directory”, and you can remove it there.
Group Access uses the same directory groups as Directory Group Entitlements.
Plugin Decision Calls
The Jev switch under TypeSafe controls plugin calls to amp.ai.decide(...) and its
single-question helpers, including calls that use Jev by default. To restrict Jev to a directory
group, turn it off on Default Policy, then grant it to that group on Group Access and save.
This does not affect Amp’s built-in use of Jev, such as ordering files in a diff. Jev cannot run an agent and does not appear in the model selector. Plugin decision calls denied by the policy fail rather than switch to OpenAI. See Decision Models.
Top Models and Personal Pins
The workspace’s Top Models, and the models members pin below them in Model Selector settings, can include a disabled model. The model selector leaves it out, and members cannot start a thread on it. The pin is kept, so it works again if you enable the model later.