Claude Code Profiles

Claude Code Profiles control how Claude Code (Anthropic CLI / IDE extension) interacts with the MoE orchestrator. A profile defines model routing, pipeline mode, and response parameters.

Concept

Claude Code communicates via the Anthropic Messages API. A profile defines:

• Which tool model (LLM for tool calls) is used
• On which endpoint (inference server) it runs
• Which MoE mode is active
• Whether and how reasoning streaming is forwarded to Claude

Profile States: enabled / disabled

Profiles no longer have an exclusive "active" flag. Instead, each profile has an enabled state (enabled: true/false):

• Multiple profiles can be enabled simultaneously
• A user must have an explicit permission grant (cc_profile) for a specific profile to access it
• The profile must also be enabled — a disabled profile is not available to any user, even if they have a permission grant
• Toggling a profile's enabled state restarts the orchestrator so the change takes effect

Orchestrator Restart

Enabling or disabling a profile triggers an orchestrator restart. Requests in progress may be interrupted.

Profile Management (/profiles)

Profile Cards

Each profile is displayed as a card:

• Enabled / Disabled badge
• Actions: Edit, Enable/Disable toggle, Delete
• Multiple profiles can be enabled at the same time

Creating / Editing a Profile

Field	Description	Example Value
Name	Unique profile name	Production – Claude Sonnet
Accepted Claude Model IDs	Comma-separated model IDs	claude-sonnet-4-6, claude-opus-4-6
Tool Model	LLM for tool calls	qwen2.5:32b or devstral:24b
Tool Endpoint	Inference server (dropdown of configured servers)	(select from Admin UI → Servers)
MoE Mode	Pipeline mode	native / moe_reasoning / moe_orchestrated
Tool Choice	How tools are invoked	auto (default) or required
Tool max_tokens	Max tokens for tool responses	8192 (min: 512)
Reasoning max_tokens	Max tokens for reasoning responses	16384 (min: 512)
System Prompt Prefix	Prepended to all system prompts	You are a helpful assistant.
Think Stream	Send MoE pipeline progress to Claude	☑

MoE Modes

Mode	Description	Latency	Quality
native	Direct to GPU, no routing	Low	Standard
moe_reasoning	Reasoning expert pipeline	Medium	High
moe_orchestrated	Full MoE pipeline (Planner → Experts → Judge)	High	Maximum

Enabling a Profile

Click the Enable toggle:

1. Sets enabled = true for this profile
2. Restarts the orchestrator (with loading overlay)
3. Users who have a permission grant for this profile can now use it

Disabling a Profile

Click the Disable toggle:

1. Sets enabled = false for this profile
2. Restarts the orchestrator
3. The profile is no longer available to any user, even if they have a permission grant

Granting Profile Access to Users

CC profiles are assigned to users via permission grants:

Admin → Edit User → Tab Permissions → Section Claude Code Profile → Select profile → +

Both conditions must be met for a user to use a profile: 1. The user has a cc_profile permission grant for the profile 2. The profile is enabled (enabled = true)

Import / Export

Export

Admin → Profiles → Export button

Downloads cc_profiles.json. See Import & Export for the full schema.

Import

Two ways to import profiles:

Option A – Upload file:

Admin → Profiles → Import JSON → Select JSON file

Option B – Paste JSON code:

Admin → Profiles → Import Code → Paste JSON → Import

Import modes:

Mode	Behavior
merge	Profiles with the same name are skipped
replace	Profiles with the same name are overwritten

Cluster Impact

Action	Effect
Create profile	Immediately available for permission grants
Edit profile	Active from the next Claude Code session
Enable profile	Orchestrator restart
Disable profile	Orchestrator restart
Delete profile	Grants for this profile are invalidated

User-Owned Profiles

Users with the expert role and a model_endpoint permission can create their own CC profiles in the User Portal. These apply only to the respective user and override the admin profile while the user is logged in.

Admins can view and manage all user profiles under /user-content.