Aller au contenu
LM Nexus centre de contrôle IA local-first
Alpha privée
FR

Configure an agent

Ce contenu n’est pas encore disponible dans votre langue.

Agent profiles let you save how an LM Nexus agent should behave, which provider/model it should use, and which tools it is allowed to see.

Open Chat, then open Chat Profiles from the profile/persona controls.

In the dialog, switch from Personas to Agents.

Choose New.

In Basics:

  1. give the agent a Name;
  2. add a short Description;
  3. choose a Default provider and model;
  4. optionally choose a Fallback provider and model;
  5. keep Agent role as Normal agent unless the profile should exist only as a delegated subagent.

Open the prompt section and describe what the agent should do, what it should prioritize, and any important limits.

Draft with assistant can propose profile text and policy, but it only edits the draft. It does not silently save the profile or execute tools.

Open Tools.

Select the Allowed Nexus tools the agent actually needs.

Then choose an Approval mode:

  • Withhold risky tools — risky tools are not exposed to the model.
  • Pause for approval — relevant tool calls can stop and ask before running.
  • Auto-run except high risk — lower-risk allowed tools can proceed while high-risk actions still pause.
  • Auto-run all allowed tools — the least restrictive profile-level mode for allowed tools.

You can also enable Auto-run read-only tools and set per-tool approval overrides.

For an Agent profile you can also configure:

  • Skills — selected skills or all skills, with load-at-start or on-demand behavior;
  • MCP server assignments — which configured MCP servers the agent can use;
  • default Notes, Knowledge, and historical-chat attachments;
  • sampling overrides;
  • agents/subagents awareness and delegation choices.

Choose Create Agent for a new profile or Save Agent for an existing one.

Back in Chat, select the saved Agent profile and the intended model/provider. Agent Runtime executes the tool calls. Chat shows the progress, tool activity, approvals, and questions.

Review the requested tool and arguments before approving it.

Approving the card resumes that exact pending action. Chat is not running a second copy of the tool call.