diff --git a/scripts/build.ts b/scripts/build.ts
index 33df301a..c799aa71 100644
--- a/scripts/build.ts
+++ b/scripts/build.ts
@@ -25,7 +25,7 @@ const featureFlags: Record<string, boolean> = {
   BRIDGE_MODE: false,
   DAEMON: false,
   AGENT_TRIGGERS: false,
-  MONITOR_TOOL: false,
+  MONITOR_TOOL: true,
   ABLATION_BASELINE: false,
   DUMP_SYSTEM_PROMPT: false,
   CACHED_MICROCOMPACT: false,
diff --git a/src/components/permissions/MonitorPermissionRequest/MonitorPermissionRequest.tsx b/src/components/permissions/MonitorPermissionRequest/MonitorPermissionRequest.tsx
new file mode 100644
index 00000000..eb8243d0
--- /dev/null
+++ b/src/components/permissions/MonitorPermissionRequest/MonitorPermissionRequest.tsx
@@ -0,0 +1,173 @@
+import React from 'react'
+import { getOriginalCwd } from '../../../bootstrap/state.js'
+import { Box, Text } from '../../../ink.js'
+import { sanitizeToolNameForAnalytics } from '../../../services/analytics/metadata.js'
+import { env } from '../../../utils/env.js'
+import { shouldShowAlwaysAllowOptions } from '../../../utils/permissions/permissionsLoader.js'
+import { usePermissionRequestLogging } from '../hooks.js'
+import { PermissionDialog } from '../PermissionDialog.js'
+import {
+  PermissionPrompt,
+  type PermissionPromptOption,
+} from '../PermissionPrompt.js'
+import type { PermissionRequestProps } from '../PermissionRequest.js'
+import { PermissionRuleExplanation } from '../PermissionRuleExplanation.js'
+import { logUnaryPermissionEvent } from '../utils.js'
+
+type OptionValue = 'yes' | 'yes-dont-ask-again' | 'no'
+
+export function MonitorPermissionRequest({
+  toolUseConfirm,
+  onDone,
+  onReject,
+  workerBadge,
+}: PermissionRequestProps) {
+  const { command, description } = toolUseConfirm.input as {
+    command?: string
+    description?: string
+  }
+
+  usePermissionRequestLogging(toolUseConfirm, {
+    completion_type: 'tool_use_single',
+    language_name: 'none',
+  })
+
+  const handleSelect = (
+    value: OptionValue,
+    feedback?: string,
+  ) => {
+    switch (value) {
+      case 'yes': {
+        logUnaryPermissionEvent({
+          completion_type: 'tool_use_single',
+          event: 'accept',
+          metadata: {
+            language_name: 'none',
+            message_id: toolUseConfirm.assistantMessage.message.id,
+            platform: env.platform,
+          },
+        })
+        toolUseConfirm.onAllow(toolUseConfirm.input, [], feedback)
+        onDone()
+        break
+      }
+      case 'yes-dont-ask-again': {
+        logUnaryPermissionEvent({
+          completion_type: 'tool_use_single',
+          event: 'accept',
+          metadata: {
+            language_name: 'none',
+            message_id: toolUseConfirm.assistantMessage.message.id,
+            platform: env.platform,
+          },
+        })
+        // Save the rule under 'Bash' toolName because checkPermissions
+        // delegates to bashToolHasPermission which matches rules against
+        // BashTool. Using 'Monitor' here would create a rule that's never
+        // checked. Command-specific prefix (like BashTool's shellRuleMatching).
+        const cmdForRule = command?.trim() || ''
+        const prefix = cmdForRule.split(/\s+/).slice(0, 2).join(' ')
+        toolUseConfirm.onAllow(toolUseConfirm.input, prefix ? [
+          {
+            type: 'addRules',
+            rules: [{ toolName: 'Bash', ruleContent: `${prefix}:*` }],
+            behavior: 'allow',
+            destination: 'localSettings',
+          },
+        ] : [])
+        onDone()
+        break
+      }
+      case 'no': {
+        logUnaryPermissionEvent({
+          completion_type: 'tool_use_single',
+          event: 'reject',
+          metadata: {
+            language_name: 'none',
+            message_id: toolUseConfirm.assistantMessage.message.id,
+            platform: env.platform,
+          },
+        })
+        toolUseConfirm.onReject(feedback)
+        onReject()
+        onDone()
+        break
+      }
+    }
+  }
+
+  const handleCancel = () => {
+    logUnaryPermissionEvent({
+      completion_type: 'tool_use_single',
+      event: 'reject',
+      metadata: {
+        language_name: 'none',
+        message_id: toolUseConfirm.assistantMessage.message.id,
+        platform: env.platform,
+      },
+    })
+    toolUseConfirm.onReject()
+    onReject()
+    onDone()
+  }
+
+  const showAlwaysAllow = shouldShowAlwaysAllowOptions()
+  const originalCwd = getOriginalCwd()
+
+  const options: PermissionPromptOption<OptionValue>[] = [
+    {
+      label: 'Yes',
+      value: 'yes',
+      feedbackConfig: { type: 'accept' },
+    },
+  ]
+
+  if (showAlwaysAllow) {
+    options.push({
+      label: (
+        <Text>
+          Yes, and don&apos;t ask again for{' '}
+          <Text bold>Monitor</Text> commands in{' '}
+          <Text bold>{originalCwd}</Text>
+        </Text>
+      ),
+      value: 'yes-dont-ask-again',
+    })
+  }
+
+  options.push({
+    label: 'No',
+    value: 'no',
+    feedbackConfig: { type: 'reject' },
+  })
+
+  const toolAnalyticsContext = {
+    toolName: sanitizeToolNameForAnalytics(toolUseConfirm.tool.name),
+    isMcp: toolUseConfirm.tool.isMcp ?? false,
+  }
+
+  return (
+    <PermissionDialog title="Monitor" workerBadge={workerBadge}>
+      <Box flexDirection="column" paddingX={2} paddingY={1}>
+        <Text>
+          Monitor({command ?? ''})
+        </Text>
+        {description ? (
+          <Text dimColor>{description}</Text>
+        ) : null}
+      </Box>
+      <Box flexDirection="column">
+        <PermissionRuleExplanation
+          permissionResult={toolUseConfirm.permissionResult}
+          toolType="tool"
+        />
+        <PermissionPrompt
+          options={options}
+          onSelect={handleSelect}
+          onCancel={handleCancel}
+          toolAnalyticsContext={toolAnalyticsContext}
+        />
+      </Box>
+    </PermissionDialog>
+  )
+}
diff --git a/src/tasks/MonitorMcpTask/MonitorMcpTask.ts b/src/tasks/MonitorMcpTask/MonitorMcpTask.ts
new file mode 100644
index 00000000..fb789a93
--- /dev/null
+++ b/src/tasks/MonitorMcpTask/MonitorMcpTask.ts
@@ -0,0 +1,102 @@
+// MonitorMcpTask — task registry entry for the 'monitor_mcp' type.
+//
+// Architecture: MonitorTool spawns shell processes as LocalShellTask
+// (type: 'local_bash', kind: 'monitor'). The 'monitor_mcp' type exists
+// in TaskType for forward-compatibility with MCP-based monitoring (not
+// yet implemented). This module satisfies the import from tasks.ts and
+// provides killMonitorMcpTasksForAgent for agent-scoped cleanup of
+// monitor-kind shell tasks.
+
+import type { AppState } from '../../state/AppState.js'
+import type { SetAppState, Task, TaskStateBase } from '../../Task.js'
+import type { AgentId } from '../../types/ids.js'
+import { logForDebugging } from '../../utils/debug.js'
+import { dequeueAllMatching } from '../../utils/messageQueueManager.js'
+import { evictTaskOutput } from '../../utils/task/diskOutput.js'
+import { updateTaskState } from '../../utils/task/framework.js'
+import { isLocalShellTask } from '../LocalShellTask/guards.js'
+import { killTask } from '../LocalShellTask/killShellTasks.js'
+
+export type MonitorMcpTaskState = TaskStateBase & {
+  type: 'monitor_mcp'
+  agentId?: AgentId
+}
+
+function isMonitorMcpTask(task: unknown): task is MonitorMcpTaskState {
+  return (
+    typeof task === 'object' &&
+    task !== null &&
+    'type' in task &&
+    task.type === 'monitor_mcp'
+  )
+}
+
+export const MonitorMcpTask: Task = {
+  name: 'MonitorMcpTask',
+  type: 'monitor_mcp',
+  async kill(taskId, setAppState) {
+    updateTaskState<MonitorMcpTaskState>(taskId, setAppState, task => {
+      if (task.status !== 'running') {
+        return task
+      }
+
+      return {
+        ...task,
+        status: 'killed',
+        notified: true,
+        endTime: Date.now(),
+      }
+    })
+    void evictTaskOutput(taskId)
+  },
+}
+
+/**
+ * Kill all monitor tasks owned by a given agent.
+ *
+ * MonitorTool spawns tasks as local_bash with kind='monitor'. When an agent
+ * exits, killShellTasksForAgent already handles those. This function provides
+ * additional cleanup for any monitor_mcp-typed tasks and also kills any
+ * local_bash tasks with kind='monitor' that might have been missed (belt and
+ * suspenders). Finally, it purges queued notifications for the dead agent.
+ */
+export function killMonitorMcpTasksForAgent(
+  agentId: AgentId,
+  getAppState: () => AppState,
+  setAppState: SetAppState,
+): void {
+  const tasks = getAppState().tasks ?? {}
+
+  for (const [taskId, task] of Object.entries(tasks)) {
+    // Kill monitor_mcp tasks for this agent
+    if (
+      isMonitorMcpTask(task) &&
+      task.agentId === agentId &&
+      task.status === 'running'
+    ) {
+      logForDebugging(
+        `killMonitorMcpTasksForAgent: killing monitor_mcp task ${taskId} (agent ${agentId} exiting)`,
+      )
+      void MonitorMcpTask.kill(taskId, setAppState)
+    }
+
+    // Also kill local_bash tasks with kind='monitor' for this agent
+    // (killShellTasksForAgent already does this, but being explicit
+    // guards against ordering issues)
+    if (
+      isLocalShellTask(task) &&
+      task.kind === 'monitor' &&
+      task.agentId === agentId &&
+      task.status === 'running'
+    ) {
+      logForDebugging(
+        `killMonitorMcpTasksForAgent: killing monitor shell task ${taskId} (agent ${agentId} exiting)`,
+      )
+      killTask(taskId, setAppState)
+    }
+  }
+
+  // Purge any queued notifications addressed to this agent — its query loop
+  // has exited and won't drain them.
+  dequeueAllMatching(cmd => cmd.agentId === agentId)
+}
diff --git a/src/tools/MonitorTool/MonitorTool.ts b/src/tools/MonitorTool/MonitorTool.ts
new file mode 100644
index 00000000..0dbb0f82
--- /dev/null
+++ b/src/tools/MonitorTool/MonitorTool.ts
@@ -0,0 +1,195 @@
+import type { ToolResultBlockParam } from '@anthropic-ai/sdk/resources/index.mjs'
+import React from 'react'
+import { z } from 'zod/v4'
+import { buildTool, type ToolDef } from '../../Tool.js'
+import { spawnShellTask } from '../../tasks/LocalShellTask/LocalShellTask.js'
+import { lazySchema } from '../../utils/lazySchema.js'
+import { exec } from '../../utils/Shell.js'
+import { getTaskOutputPath } from '../../utils/task/diskOutput.js'
+import {
+  bashToolHasPermission,
+  matchWildcardPattern,
+  permissionRuleExtractPrefix,
+} from '../BashTool/bashPermissions.js'
+import { parseForSecurity } from '../../utils/bash/ast.js'
+
+export const MONITOR_TOOL_NAME = 'Monitor'
+
+const MONITOR_TIMEOUT_MS = 30 * 60 * 1000 // 30 minutes
+
+const inputSchema = lazySchema(() =>
+  z.strictObject({
+    command: z
+      .string()
+      .describe('The shell command to run and monitor'),
+    description: z
+      .string()
+      .describe(
+        'Clear, concise description of what this command does in active voice.',
+      ),
+  }),
+)
+type InputSchema = ReturnType<typeof inputSchema>
+
+const outputSchema = lazySchema(() =>
+  z.object({
+    taskId: z
+      .string()
+      .describe('The ID of the background monitor task'),
+    outputFile: z
+      .string()
+      .describe('Path to the file where output is being written'),
+  }),
+)
+type OutputSchema = ReturnType<typeof outputSchema>
+
+type Output = z.infer<OutputSchema>
+
+export const MonitorTool = buildTool({
+  name: MONITOR_TOOL_NAME,
+  searchHint: 'stream shell output as notifications',
+  maxResultSizeChars: 10_000,
+  strict: true,
+
+  isConcurrencySafe() {
+    return true
+  },
+
+  toAutoClassifierInput(input) {
+    return input.command
+  },
+
+  async preparePermissionMatcher({ command }) {
+    const parsed = await parseForSecurity(command)
+    if (parsed.kind !== 'simple') {
+      return () => true
+    }
+    const subcommands = parsed.commands.map(c => c.argv.join(' '))
+    return (pattern: string) => {
+      const prefix = permissionRuleExtractPrefix(pattern)
+      return subcommands.some(cmd => {
+        if (prefix !== null) {
+          return cmd === prefix || cmd.startsWith(`${prefix} `)
+        }
+        return matchWildcardPattern(pattern, cmd)
+      })
+    }
+  },
+
+  async checkPermissions(input, context) {
+    // Delegate to the bash permission system — Monitor runs shell commands
+    // just like Bash does, so the same permission rules apply.
+    return bashToolHasPermission({ command: input.command }, context)
+  },
+
+  async description(input) {
+    return input.description || 'Monitor shell command'
+  },
+
+  async prompt() {
+    return `Execute a shell command in the background and stream its stdout line-by-line as notifications. Each polling interval (~1s), new output lines are delivered to you. Use this for monitoring logs, watching build output, or observing long-running processes. For one-shot "wait until done" commands, prefer Bash with run_in_background instead.`
+  },
+
+  get inputSchema(): InputSchema {
+    return inputSchema()
+  },
+
+  get outputSchema(): OutputSchema {
+    return outputSchema()
+  },
+
+  userFacingName() {
+    return 'Monitor'
+  },
+
+  getToolUseSummary(input) {
+    if (!input?.description) {
+      return input?.command ?? null
+    }
+    return input.description
+  },
+
+  getActivityDescription(input) {
+    if (!input?.description) {
+      return 'Starting monitor'
+    }
+    return `Monitoring ${input.description}`
+  },
+
+  renderToolUseMessage(
+    input: Partial<z.infer<InputSchema>>,
+  ): React.ReactNode {
+    const cmd = input.command ?? ''
+    const desc = input.description ?? ''
+    if (desc && cmd) {
+      return `${desc}: ${cmd}`
+    }
+    return cmd || desc || ''
+  },
+
+  renderToolResultMessage(
+    output: Output,
+  ): React.ReactNode {
+    return `Monitor started (task ${output.taskId})`
+  },
+
+  mapToolResultToToolResultBlockParam(
+    output: Output,
+    toolUseID: string,
+  ): ToolResultBlockParam {
+    const outputPath = output.outputFile
+    return {
+      tool_use_id: toolUseID,
+      type: 'tool_result',
+      content: `Monitor task started with ID: ${output.taskId}. Output is being streamed to: ${outputPath}. You will receive notifications as new output lines appear (~1s polling). Use TaskStop to end monitoring when done.`,
+    }
+  },
+
+  async call(input, toolUseContext) {
+    const { command, description } = input
+    const { abortController, setAppState } = toolUseContext
+
+    // Create the shell command — uses the same Shell.exec() as BashTool.
+    // This is intentionally a shell execution (not execFile) because
+    // MonitorTool needs full shell features (pipes, redirects, etc.)
+    // just like BashTool does.
+    const shellCommand = await exec(
+      command,
+      abortController.signal,
+      'bash',
+      { timeout: MONITOR_TIMEOUT_MS },
+    )
+
+    // Spawn as a background task with kind='monitor' — identical to
+    // BashTool's run_in_background path but always monitor-flavored.
+    const handle = await spawnShellTask(
+      {
+        command,
+        description: description || command,
+        shellCommand,
+        toolUseId: toolUseContext.toolUseId,
+        agentId: toolUseContext.agentId,
+        kind: 'monitor',
+      },
+      {
+        abortController,
+        getAppState: () => {
+          throw new Error(
+            'getAppState not available in MonitorTool spawn context',
+          )
+        },
+        setAppState: toolUseContext.setAppStateForTasks ?? setAppState,
+      },
+    )
+
+    const taskId = handle.taskId
+    const outputFile = getTaskOutputPath(taskId)
+
+    return {
+      data: {
+        taskId,
+        outputFile,
+      },
+    }
+  },
+} satisfies ToolDef<InputSchema, Output>)