zed/docs/src/ai/agent-settings.md

title, description

title	description
Agent Settings - Zed	Map the AI settings pages to Zed AI setup for LLM providers, External Agents, MCP servers, and related settings.

Agent Settings

Agent Settings live in the AI section of the Settings Editor, which configures model providers, External Agents, and MCP servers. Open it with {#action agent::OpenSettings} (also available from the top-right menu in the Agent Panel), which takes you straight to the AI page. You can also reach the same page with {#action zed::OpenSettings} and selecting AI in the sidebar.

Within the AI page, LLM Providers, External Agents, and MCP Servers each open as their own sub-page under the General section.

Surface	Opens with	Use it for
Settings Editor	{#action agent::OpenSettings}	AI settings: LLM providers, External Agents, MCP servers, and related pages
Settings file	{#action zed::OpenSettingsFile}	Direct JSON edits and settings not exposed in UI

For general settings mechanics, see Configuring Zed.

LLM Providers

The LLM Providers section configures model providers for Zed AI features, including Zed Agent, Inline Assistant, Git commit generation, thread summaries, and similar model-backed features.

Use this section to:

• sign in to supported subscription-backed providers
• enter provider API keys
• add OpenAI-compatible providers
• remove providers

For the model-access paths and provider-specific setup, see LLM Providers.

Feature-Specific Settings

Some Zed AI features have their own model or prompt settings in settings.json, including:

• agent.inline_assistant_model
• agent.commit_message_model
• agent.thread_summary_model
• agent.compaction_model
• agent.subagent_model
• agent.commit_message_instructions
• agent.inline_alternatives

Use agent.commit_message_instructions for instructions that apply only to generated Git commit messages:

{
  "agent": {
    "commit_message_instructions": "Use the Conventional Commits format: <type>(<scope>): <description>."
  }
}

For feature-specific model examples, see Feature-specific Models.

Automatic Compaction

Zed Agent can automatically compact long threads before they reach the selected model's context window. Compaction summarizes earlier messages and keeps the conversation usable without starting a new thread.

Automatic compaction is enabled by default and runs when the thread reaches 90% of the model's context window. You can change the threshold or disable automatic compaction in settings.json:

{
  "agent": {
    "auto_compact": {
      "enabled": true,
      "threshold": "90%"
    }
  }
}

The threshold value can be one of:

Value	Meaning
Percentage string, like 90%	Compact when the thread uses that percentage of the model's context window.
Positive integer, like 100000	Compact after that many tokens have been used.
Negative integer, like -20000	Compact once fewer than that many tokens remain in the model's context window.

0 is not a valid threshold. If the threshold is invalid, Zed falls back to 90%.

You can compact a Zed Agent thread manually at any time by typing /compact in the Agent Panel message editor. For more on thread token usage and compaction behavior, see Token Usage and Compaction.

Compaction Model

By default, context compaction (both /compact and auto-compaction) uses the thread's currently selected model. Set agent.compaction_model to use a different model:

{
  "agent": {
    "default_model": {
      "provider": "anthropic",
      "model": "claude-opus-4-6"
    },
    "compaction_model": {
      "provider": "anthropic",
      "model": "claude-sonnet-4-5"
    }
  }
}

Notes:

• The configured model should have a context window at least as large as the thread's primary model for predictable behavior.

External Agents

The External Agents section configures ACP-integrated agents.

Use Add Agent to:

• Install from Registry
• Add Custom Agent

For setup details and support boundaries, see External Agents.

MCP Servers

The MCP Servers page configures Model Context Protocol servers connected to Zed.

Use Add Server (in the page header) to:

• Add Local Server
• Add Remote Server
• Install from Extensions

Each configured server row exposes:

• a Configure (gear) button to edit the server
• an Uninstall (trash) button to remove it
• a toggle to enable or disable it

For MCP setup, auth, server status, and agent-path boundaries, see MCP.

Related Configuration

Some AI settings are not configured on the AI settings pages:

Task	Go to
Choose which tools are available in a Zed Agent thread	Agent Profiles
Control whether tool calls are allowed, denied, or confirmed	Tool Permissions
Configure reusable task instructions	Skills
Configure always-on personal or project instructions	Instructions
Configure edit prediction providers	Edit Prediction
Turn AI off	AI Quick Start
Edit raw settings JSON	Configuring Zed

Feature-Specific Models

Zed supports feature-specific model settings for Inline Assistant, Git commit generation, thread summaries, and subagents. Configure these in settings when you need a different model for a specific workflow.

See LLM Providers for model access, and All Settings for the complete settings reference.

Model Temperature

Most Zed AI features use the selected model's default generation behavior. Use agent.model_parameters when you need to set a temperature for a provider, a model, or a specific provider/model pair.

{
  "agent": {
    "model_parameters": [
      {
        "provider": "anthropic",
        "model": "claude-sonnet-4-5",
        "temperature": 0.2
      }
    ]
  }
}

Zed checks matching entries from last to first. An entry can omit provider or model to apply more broadly. For provider-specific model configuration such as custom model entries, context windows, or gateway routing, see LLM Providers and the provider setup pages.

Rules, Skills, and Instructions

Reusable Rules have been replaced by Skills. Always-on Rules have moved to Instructions, including personal AGENTS.md and project instruction files.

Older builds or transitional UI may still refer to Rules. Use Skills for reusable task instructions and Instructions for always-on context.