Agent Settings

Settings for Zed's Agent Panel, including model selection, UI preferences, and tool permissions.

If you're using Zed's hosted LLM service, it sets claude-sonnet-4-5 as the default model for agentic work (agent panel, inline assistant) and gpt-5-nano as the default "fast" model (thread summarization, git commit messages). If you're not subscribed or want to change these defaults, you can manually edit the default_model object in your settings:

{
  "agent": {
    "default_model": {
      "provider": "openai",
      "model": "gpt-4o"
    }
  }
}

Feature-specific Models

You can assign distinct and specific models for the following AI-powered features:

Thread summary model: Used for generating thread summaries
Inline assistant model: Used for the inline assistant feature
Commit message model: Used for generating Git commit messages

{
  "agent": {
    "default_model": {
      "provider": "zed.dev",
      "model": "claude-sonnet-4-5"
    },
    "inline_assistant_model": {
      "provider": "anthropic",
      "model": "claude-3-5-sonnet"
    },
    "commit_message_model": {
      "provider": "openai",
      "model": "gpt-4o-mini"
    },
    "thread_summary_model": {
      "provider": "google",
      "model": "gemini-2.0-flash"
    }
  }
}

If a custom model isn't set for one of these features, they automatically fall back to using the default model.

Alternative Models for Inline Assists

With the Inline Assistant in particular, you can send the same prompt to multiple models at once.

Here's how you can customize your settings file (how to edit) to add this functionality:

{
  "agent": {
    "default_model": {
      "provider": "zed.dev",
      "model": "claude-sonnet-4-5"
    },
    "inline_alternatives": [
      {
        "provider": "zed.dev",
        "model": "gpt-5-mini"
      }
    ]
  }
}

When multiple models are configured, you'll see in the Inline Assistant UI buttons that allow you to cycle between outputs generated by each model.

The models you specify here are always used in addition to your default model.

For example, the following configuration will generate three outputs for every assist. One with Claude Sonnet 4.5 (the default model), another with GPT-5-mini, and another one with Gemini 3 Flash.

{
  "agent": {
    "default_model": {
      "provider": "zed.dev",
      "model": "claude-sonnet-4-5"
    },
    "inline_alternatives": [
      {
        "provider": "zed.dev",
        "model": "gpt-5-mini"
      },
      {
        "provider": "zed.dev",
        "model": "gemini-3-flash"
      }
    ]
  }
}

Model Temperature

Specify a custom temperature for a provider and/or model:

{
  "agent": {
    "model_parameters": [
      // To set parameters for all requests to OpenAI models:
      {
        "provider": "openai",
        "temperature": 0.5
      },
      // To set parameters for all requests in general:
      {
        "temperature": 0
      },
      // To set parameters for a specific provider and model:
      {
        "provider": "zed.dev",
        "model": "claude-sonnet-4-5",
        "temperature": 1.0
      }
    ]
  }
}

Agent Panel Settings

Note that some of these settings are also surfaced in the Agent Panel's settings UI, which you can access either via the agent: open settings action or by the dropdown menu on the top-right corner of the panel.

Default View

Use the default_view setting to change the default view of the Agent Panel. You can choose between thread (the default) and text_thread:

{
  "agent": {
    "default_view": "text_thread"
  }
}

Default Tool Permissions

Note: In Zed v0.224.0 and above, tool approval uses the agent.tool_permissions settings described below.

The agent.tool_permissions.default setting controls the baseline tool approval behavior for Zed's native agent:

"confirm" (default) — Prompts for approval before running any tool action
"allow" — Auto-approves tool actions without prompting
"deny" — Blocks all tool actions

{
  "agent": {
    "tool_permissions": {
      "default": "confirm"
    }
  }
}

Even with "default": "allow", per-tool always_deny and always_confirm patterns are still respected, so you can auto-approve most actions while keeping guardrails on dangerous or sensitive ones.

Per-tool Permission Rules

For granular control over individual tool actions, use the tools key inside tool_permissions to configure regex-based rules that auto-approve, auto-deny, or always require confirmation for specific inputs.

Each tool entry supports the following keys:

default — Fallback when no patterns match: "confirm", "allow", or "deny"
always_allow — Array of patterns that auto-approve matching actions
always_deny — Array of patterns that block matching actions immediately
always_confirm — Array of patterns that always prompt for confirmation

{
  "agent": {
    "tool_permissions": {
      "default": "allow",
      "tools": {
        "terminal": {
          "default": "confirm",
          "always_allow": [
            { "pattern": "^cargo\\s+(build|test|check)" },
            { "pattern": "^git\\s+(status|log|diff)" }
          ],
          "always_deny": [{ "pattern": "rm\\s+-rf\\s+(/|~)" }],
          "always_confirm": [{ "pattern": "sudo\\s" }]
        },
        "edit_file": {
          "always_deny": [
            { "pattern": "\\.env" },
            { "pattern": "\\.(pem|key)$" }
          ]
        }
      }
    }
  }
}

Pattern Precedence

When evaluating a tool action, rules are checked in the following order (highest priority first):

Built-in security rules — Hardcoded protections (e.g., rm -rf /) that cannot be overridden
always_deny — Blocks matching actions immediately
always_confirm — Requires confirmation for matching actions
always_allow — Auto-approves matching actions. For the terminal tool with chained commands (e.g., echo hello && rm file), all sub-commands must match an always_allow pattern
Tool-specific default — Per-tool fallback when no patterns match (e.g., tools.terminal.default)
Global default — Falls back to tool_permissions.default

Case Sensitivity

Patterns are case-insensitive by default. To make a pattern case-sensitive, set case_sensitive to true:

{
  "agent": {
    "tool_permissions": {
      "tools": {
        "edit_file": {
          "always_deny": [
            {
              "pattern": "^Makefile$",
              "case_sensitive": true
            }
          ]
        }
      }
    }
  }
}

`copy_path` and `move_path` Patterns

For the copy_path and move_path tools, patterns are matched independently against both the source and destination paths. A deny or confirm match on either path takes effect. For always_allow, both paths must match for auto-approval.

MCP Tool Permissions

MCP tools use the key format mcp:<server_name>:<tool_name> in the tools configuration. For example:

{
  "agent": {
    "tool_permissions": {
      "tools": {
        "mcp:github:create_issue": {
          "default": "confirm"
        },
        "mcp:github:create_pull_request": {
          "default": "deny"
        }
      }
    }
  }
}

The default key on each MCP tool entry is the primary mechanism for controlling MCP tool permissions. Pattern-based rules (always_allow, always_deny, always_confirm) match against an empty string for MCP tools, so most patterns won't match — use the tool-level default instead.

See the Tool Permissions documentation for more examples and complete details.

Note: Before Zed v0.224.0, tool approval was controlled by the agent.always_allow_tool_actions boolean (default false). Set it to true to auto-approve tool actions, or leave it false to require confirmation for edits and tool calls.

Single-file Review

Control whether to display review actions (accept & reject) in single buffers after the agent is done performing edits. The default value is true.

{
  "agent": {
    "single_file_review": false
  }
}

When set to false, these controls are only available in the multibuffer review tab.

Sound Notification

Control whether to hear a notification sound when the agent is done generating changes or needs your input. The default value is false.

{
  "agent": {
    "play_sound_when_agent_done": true
  }
}

Message Editor Size

Use the message_editor_min_lines setting to control the minimum number of lines of height the agent message editor should have. It is set to 4 by default, and the max number of lines is always double of the minimum.

{
  "agent": {
    "message_editor_min_lines": 4
  }
}

Modifier to Send

Require a modifier (cmd on macOS, ctrl on Linux) to send messages. Prevents accidental sends while editing. The default value is false.

{
  "agent": {
    "use_modifier_to_send": true
  }
}

Edit Card

Use the expand_edit_card setting to control whether edit cards show the full diff in the Agent Panel. It is set to true by default, but if set to false, the card's height is capped to a certain number of lines, requiring a click to be expanded.

{
  "agent": {
    "expand_edit_card": false
  }
}

Terminal Card

Use the expand_terminal_card setting to control whether terminal cards show the command output in the Agent Panel. It is set to true by default, but if set to false, the card will be fully collapsed even while the command is running, requiring a click to be expanded.

{
  "agent": {
    "expand_terminal_card": false
  }
}

Feedback Controls

Control whether to display the thumbs up/down buttons at the bottom of each agent response, allowing you to give Zed feedback about the agent's performance. The default value is true.

{
  "agent": {
    "enable_feedback": false
  }
}