diff --git a/packages/api/src/agents/hitl/index.ts b/packages/api/src/agents/hitl/index.ts
new file mode 100644
index 0000000000..62d5b849e4
--- /dev/null
+++ b/packages/api/src/agents/hitl/index.ts
@@ -0,0 +1 @@
+export * from './policy';
diff --git a/packages/api/src/agents/hitl/policy.spec.ts b/packages/api/src/agents/hitl/policy.spec.ts
new file mode 100644
index 0000000000..2387d77f0f
--- /dev/null
+++ b/packages/api/src/agents/hitl/policy.spec.ts
@@ -0,0 +1,125 @@
+import type { TToolApprovalPolicy } from 'librechat-data-provider';
+import { decideToolApproval, requiresApproval, buildPendingAction } from './policy';
+
+describe('decideToolApproval', () => {
+  test('returns "allow" when no policy is configured', () => {
+    expect(decideToolApproval(undefined, { name: 'shell' })).toBe('allow');
+  });
+
+  test('returns the configured default when no rule matches', () => {
+    const policy: TToolApprovalPolicy = { default: 'ask' };
+    expect(decideToolApproval(policy, { name: 'unmapped' })).toBe('ask');
+  });
+
+  test('falls back to "allow" when default is omitted', () => {
+    const policy: TToolApprovalPolicy = { required: ['shell'] };
+    expect(decideToolApproval(policy, { name: 'web_search' })).toBe('allow');
+  });
+
+  test('returns "ask" when tool is in required list', () => {
+    const policy: TToolApprovalPolicy = { required: ['shell', 'execute_code'] };
+    expect(decideToolApproval(policy, { name: 'execute_code' })).toBe('ask');
+  });
+
+  test('returns "allow" for tools in excluded list, even when default is "ask"', () => {
+    const policy: TToolApprovalPolicy = { default: 'ask', excluded: ['web_search'] };
+    expect(decideToolApproval(policy, { name: 'web_search' })).toBe('allow');
+  });
+
+  test('excluded wins over required when a tool appears in both (defensive)', () => {
+    const policy: TToolApprovalPolicy = {
+      default: 'allow',
+      required: ['shell'],
+      excluded: ['shell'],
+    };
+    expect(decideToolApproval(policy, { name: 'shell' })).toBe('allow');
+  });
+
+  test('returns the default when tool name is missing', () => {
+    const policy: TToolApprovalPolicy = { default: 'ask', required: ['shell'] };
+    expect(decideToolApproval(policy, {})).toBe('ask');
+  });
+});
+
+describe('requiresApproval', () => {
+  test('true only when decision is "ask"', () => {
+    const policy: TToolApprovalPolicy = { required: ['shell'] };
+    expect(requiresApproval(policy, { name: 'shell' })).toBe(true);
+    expect(requiresApproval(policy, { name: 'web_search' })).toBe(false);
+    expect(requiresApproval(undefined, { name: 'shell' })).toBe(false);
+  });
+});
+
+describe('buildPendingAction', () => {
+  const baseInput = {
+    streamId: 'stream-1',
+    conversationId: 'conv-1',
+    runId: 'run-1',
+    responseMessageId: 'msg-1',
+    toolCalls: [
+      {
+        name: 'shell',
+        arguments: { command: 'ls' },
+        tool_call_id: 'call_abc',
+        description: 'List files',
+      },
+    ],
+  };
+
+  test('produces a payload mirroring LangChain HumanInterrupt shape', () => {
+    const action = buildPendingAction(baseInput);
+    expect(action.payload.type).toBe('tool_approval');
+    expect(action.payload.action_requests).toEqual([
+      {
+        name: 'shell',
+        arguments: { command: 'ls' },
+        tool_call_id: 'call_abc',
+        description: 'List files',
+      },
+    ]);
+    expect(action.payload.review_configs).toEqual([
+      { action_name: 'shell', allowed_decisions: ['approve', 'reject', 'edit'] },
+    ]);
+  });
+
+  test('respects per-tool decision overrides', () => {
+    const action = buildPendingAction({
+      ...baseInput,
+      decisionsByToolName: { shell: ['approve', 'reject'] },
+    });
+    expect(action.payload.review_configs[0].allowed_decisions).toEqual(['approve', 'reject']);
+  });
+
+  test('generates a uuid actionId by default', () => {
+    const a = buildPendingAction(baseInput);
+    const b = buildPendingAction(baseInput);
+    expect(a.actionId).not.toBe(b.actionId);
+    expect(a.actionId).toMatch(/^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i);
+  });
+
+  test('honours an explicit actionId', () => {
+    const action = buildPendingAction({ ...baseInput, actionId: 'fixed-id' });
+    expect(action.actionId).toBe('fixed-id');
+  });
+
+  test('sets expiresAt only when ttlMs is provided', () => {
+    const without = buildPendingAction(baseInput);
+    expect(without.expiresAt).toBeUndefined();
+
+    const ttl = 5_000;
+    const before = Date.now();
+    const withTtl = buildPendingAction({ ...baseInput, ttlMs: ttl });
+    const after = Date.now();
+    expect(withTtl.expiresAt).toBeDefined();
+    expect(withTtl.expiresAt).toBeGreaterThanOrEqual(before + ttl);
+    expect(withTtl.expiresAt).toBeLessThanOrEqual(after + ttl);
+  });
+
+  test('preserves stream/conversation/run identifiers', () => {
+    const action = buildPendingAction(baseInput);
+    expect(action.streamId).toBe('stream-1');
+    expect(action.conversationId).toBe('conv-1');
+    expect(action.runId).toBe('run-1');
+    expect(action.responseMessageId).toBe('msg-1');
+  });
+});
diff --git a/packages/api/src/agents/hitl/policy.ts b/packages/api/src/agents/hitl/policy.ts
new file mode 100644
index 0000000000..9f6516ce39
--- /dev/null
+++ b/packages/api/src/agents/hitl/policy.ts
@@ -0,0 +1,113 @@
+import { randomUUID } from 'crypto';
+import type { Agents, TToolApprovalPolicy, ToolApprovalDecision } from 'librechat-data-provider';
+
+/**
+ * Default decision set offered to the user for a paused tool call.
+ *
+ * `approve` runs the tool as-is, `reject` blocks it with a rejection message,
+ * `edit` re-runs it with user-supplied arguments. `respond` is reserved for
+ * the future ask-user-question flow and intentionally NOT in the default set.
+ */
+const DEFAULT_REVIEW_DECISIONS: Agents.ToolApprovalDecision[] = ['approve', 'reject', 'edit'];
+
+/**
+ * Tool reference accepted by the policy resolver.
+ * Loosened from `Agents.ToolCall` so callers can pass minimal shapes
+ * (e.g. SDK hook payloads, MCP-derived names) without re-typing.
+ */
+export interface ToolRef {
+  name?: string;
+}
+
+/**
+ * Decide whether a tool call requires human approval, denial, or can run as-is.
+ *
+ * Resolution order:
+ *   1. `excluded` (always allow) wins over everything else.
+ *   2. `required` (always ask) wins over the default.
+ *   3. Falls back to `policy.default`, which itself defaults to `'allow'`.
+ *
+ * Returns `'allow'` when no policy is configured or the tool name is missing.
+ */
+export function decideToolApproval(
+  policy: TToolApprovalPolicy | undefined,
+  tool: ToolRef,
+): ToolApprovalDecision {
+  if (!policy) {
+    return 'allow';
+  }
+  const name = tool.name;
+  const fallback = policy.default ?? 'allow';
+  if (!name) {
+    return fallback;
+  }
+  if (policy.excluded?.includes(name)) {
+    return 'allow';
+  }
+  if (policy.required?.includes(name)) {
+    return 'ask';
+  }
+  return fallback;
+}
+
+/** Convenience wrapper. True when the tool call should be paused for human review. */
+export function requiresApproval(policy: TToolApprovalPolicy | undefined, tool: ToolRef): boolean {
+  return decideToolApproval(policy, tool) === 'ask';
+}
+
+/** Input shape for {@link buildPendingAction}. */
+export interface BuildPendingActionInput {
+  streamId: string;
+  conversationId?: string;
+  /** Stable per-turn identifier (e.g. responseMessageId or LangGraph checkpoint_ns). */
+  runId?: string;
+  responseMessageId?: string;
+  /** One entry per tool call awaiting review in this interrupt. */
+  toolCalls: Array<{
+    name: string;
+    arguments: string | Record<string, unknown>;
+    tool_call_id: string;
+    description?: string;
+  }>;
+  /** Override decisions per tool name. Falls back to {@link DEFAULT_REVIEW_DECISIONS}. */
+  decisionsByToolName?: Record<string, Agents.ToolApprovalDecision[]>;
+  /** Optional TTL (ms). When set, `expiresAt = createdAt + ttlMs`. */
+  ttlMs?: number;
+  /** Override actionId; defaults to a fresh uuid. */
+  actionId?: string;
+}
+
+/**
+ * Build a {@link Agents.PendingAction} record from one or more paused tool calls.
+ *
+ * The resulting `payload` mirrors LangChain HITL middleware's `HumanInterrupt` shape
+ * (`action_requests` + `review_configs`) so it can be forwarded directly when the
+ * SDK adopts native HITL primitives.
+ */
+export function buildPendingAction(input: BuildPendingActionInput): Agents.PendingAction {
+  const createdAt = Date.now();
+  const action_requests: Agents.ToolApprovalRequest[] = input.toolCalls.map((tc) => ({
+    name: tc.name,
+    arguments: tc.arguments,
+    tool_call_id: tc.tool_call_id,
+    description: tc.description,
+  }));
+  const review_configs: Agents.ToolReviewConfig[] = input.toolCalls.map((tc) => ({
+    action_name: tc.name,
+    allowed_decisions: input.decisionsByToolName?.[tc.name] ?? DEFAULT_REVIEW_DECISIONS,
+  }));
+  return {
+    actionId: input.actionId ?? randomUUID(),
+    streamId: input.streamId,
+    conversationId: input.conversationId,
+    runId: input.runId,
+    responseMessageId: input.responseMessageId,
+    payload: {
+      type: 'tool_approval',
+      action_requests,
+      review_configs,
+    },
+    createdAt,
+    expiresAt: input.ttlMs ? createdAt + input.ttlMs : undefined,
+  };
+}
diff --git a/packages/api/src/agents/index.ts b/packages/api/src/agents/index.ts
index 14e0f480fd..fc14367e3e 100644
--- a/packages/api/src/agents/index.ts
+++ b/packages/api/src/agents/index.ts
@@ -27,3 +27,4 @@ export * from './tools';
 export * from './validation';
 export * from './added';
 export * from './load';
+export * from './hitl';
diff --git a/packages/api/src/stream/GenerationJobManager.ts b/packages/api/src/stream/GenerationJobManager.ts
index 238db4e6cb..3ae476b245 100644
--- a/packages/api/src/stream/GenerationJobManager.ts
+++ b/packages/api/src/stream/GenerationJobManager.ts
@@ -1313,6 +1313,54 @@ class GenerationJobManagerClass {
     this.jobStore.setGraph(streamId, graph);
   }
 
+  /**
+   * Transition a job to `requires_action` and persist the pending review record.
+   *
+   * The job is NOT cleaned up: chunks, run steps, and user-active-set membership
+   * remain so the resume path can rebuild context. The Redis job-hash TTL is
+   * refreshed by the store to give the user the full TTL window to respond.
+   *
+   * @param streamId - The stream identifier
+   * @param pendingAction - The pending review record (tool approval, etc.)
+   */
+  async markRequiresAction(streamId: string, pendingAction: Agents.PendingAction): Promise<void> {
+    await this.jobStore.updateJob(streamId, {
+      status: 'requires_action',
+      pendingAction,
+    });
+    logger.debug(
+      `[GenerationJobManager] Job awaiting human review: ${streamId} action=${pendingAction.actionId}`,
+    );
+  }
+
+  /**
+   * Read the pending review record for a job.
+   *
+   * Returns null when the job doesn't exist, isn't in `requires_action`,
+   * or has no recorded pending action. Callers (status endpoint, approval routes)
+   * should treat null as "nothing to approve."
+   */
+  async getPendingAction(streamId: string): Promise<Agents.PendingAction | null> {
+    const jobData = await this.jobStore.getJob(streamId);
+    if (!jobData || jobData.status !== 'requires_action') {
+      return null;
+    }
+    return jobData.pendingAction ?? null;
+  }
+
+  /**
+   * Clear the pending review record and return the job to `running`.
+   * Called by the resume path after a user approval/rejection has been accepted
+   * and the run is about to be re-driven.
+   */
+  async clearPendingAction(streamId: string): Promise<void> {
+    await this.jobStore.updateJob(streamId, {
+      status: 'running',
+      pendingAction: undefined,
+    });
+    logger.debug(`[GenerationJobManager] Cleared pending action: ${streamId}`);
+  }
+
   /**
    * Get resume state for reconnecting clients.
    */
@@ -1538,13 +1586,14 @@ class GenerationJobManagerClass {
    * Get job count by status.
    */
   async getJobCountByStatus(): Promise<Record<t.GenerationJobStatus, number>> {
-    const [running, complete, error, aborted] = await Promise.all([
+    const [running, complete, error, aborted, requires_action] = await Promise.all([
       this.jobStore.getJobCountByStatus('running'),
       this.jobStore.getJobCountByStatus('complete'),
       this.jobStore.getJobCountByStatus('error'),
       this.jobStore.getJobCountByStatus('aborted'),
+      this.jobStore.getJobCountByStatus('requires_action'),
     ]);
-    return { running, complete, error, aborted };
+    return { running, complete, error, aborted, requires_action };
   }
 
   /**
diff --git a/packages/api/src/stream/__tests__/pendingAction.spec.ts b/packages/api/src/stream/__tests__/pendingAction.spec.ts
new file mode 100644
index 0000000000..11dc991ed9
--- /dev/null
+++ b/packages/api/src/stream/__tests__/pendingAction.spec.ts
@@ -0,0 +1,107 @@
+import type { Agents } from 'librechat-data-provider';
+import { InMemoryEventTransport } from '~/stream/implementations/InMemoryEventTransport';
+import { InMemoryJobStore } from '~/stream/implementations/InMemoryJobStore';
+import { GenerationJobManagerClass } from '~/stream/GenerationJobManager';
+import { buildPendingAction } from '~/agents/hitl/policy';
+
+jest.spyOn(console, 'log').mockImplementation();
+
+describe('GenerationJobManager pending-action lifecycle (in-memory)', () => {
+  let manager: GenerationJobManagerClass;
+
+  beforeEach(() => {
+    manager = new GenerationJobManagerClass();
+    manager.configure({
+      jobStore: new InMemoryJobStore({ ttlAfterComplete: 60000 }),
+      eventTransport: new InMemoryEventTransport(),
+      isRedis: false,
+      cleanupOnComplete: false,
+    });
+    manager.initialize();
+  });
+
+  afterEach(async () => {
+    await manager.destroy();
+  });
+
+  function buildAction(streamId: string, overrides: Partial<Agents.PendingAction> = {}) {
+    const action = buildPendingAction({
+      streamId,
+      conversationId: streamId,
+      runId: 'run-1',
+      responseMessageId: 'msg-1',
+      toolCalls: [{ name: 'shell', arguments: { command: 'ls' }, tool_call_id: 'call_abc' }],
+    });
+    return { ...action, ...overrides };
+  }
+
+  test('markRequiresAction persists the pending action and transitions status', async () => {
+    const streamId = 'stream-mark';
+    await manager.createJob(streamId, 'user-1');
+
+    const action = buildAction(streamId);
+    await manager.markRequiresAction(streamId, action);
+
+    const status = await manager.getJobStatus(streamId);
+    expect(status).toBe('requires_action');
+
+    const pending = await manager.getPendingAction(streamId);
+    expect(pending).not.toBeNull();
+    expect(pending?.actionId).toBe(action.actionId);
+    expect(pending?.payload.action_requests[0].name).toBe('shell');
+  });
+
+  test('getPendingAction returns null for jobs not in requires_action', async () => {
+    const streamId = 'stream-running';
+    await manager.createJob(streamId, 'user-1');
+    expect(await manager.getPendingAction(streamId)).toBeNull();
+  });
+
+  test('getPendingAction returns null when the job does not exist', async () => {
+    expect(await manager.getPendingAction('nonexistent')).toBeNull();
+  });
+
+  test('clearPendingAction returns the job to running and removes the pending record', async () => {
+    const streamId = 'stream-clear';
+    await manager.createJob(streamId, 'user-1');
+
+    await manager.markRequiresAction(streamId, buildAction(streamId));
+    expect(await manager.getJobStatus(streamId)).toBe('requires_action');
+
+    await manager.clearPendingAction(streamId);
+
+    expect(await manager.getJobStatus(streamId)).toBe('running');
+    expect(await manager.getPendingAction(streamId)).toBeNull();
+  });
+
+  test('requires_action drops the running count but keeps the user-active set', async () => {
+    const streamId = 'stream-counts';
+    await manager.createJob(streamId, 'user-counts');
+
+    const beforeCounts = await manager.getJobCountByStatus();
+    expect(beforeCounts.running).toBe(1);
+    expect(beforeCounts.requires_action).toBe(0);
+
+    await manager.markRequiresAction(streamId, buildAction(streamId));
+
+    const afterCounts = await manager.getJobCountByStatus();
+    expect(afterCounts.running).toBe(0);
+    expect(afterCounts.requires_action).toBe(1);
+
+    // Pending-approval jobs still occupy the user's conversation slot.
+    const active = await manager.getActiveJobIdsForUser('user-counts');
+    expect(active).toContain(streamId);
+  });
+
+  test('getActiveJobIdsForUser excludes terminal jobs but includes requires_action', async () => {
+    await manager.createJob('s-running', 'user-mix');
+    await manager.createJob('s-paused', 'user-mix');
+    await manager.createJob('s-done', 'user-mix');
+
+    await manager.markRequiresAction('s-paused', buildAction('s-paused'));
+    await manager.completeJob('s-done');
+
+    const active = await manager.getActiveJobIdsForUser('user-mix');
+    expect(active.sort()).toEqual(['s-paused', 's-running']);
+  });
+});
diff --git a/packages/api/src/stream/implementations/InMemoryJobStore.ts b/packages/api/src/stream/implementations/InMemoryJobStore.ts
index f318f3baf5..4567b55ae6 100644
--- a/packages/api/src/stream/implementations/InMemoryJobStore.ts
+++ b/packages/api/src/stream/implementations/InMemoryJobStore.ts
@@ -295,8 +295,9 @@ export class InMemoryJobStore implements IJobStore {
 
     for (const streamId of trackedIds) {
       const job = this.jobs.get(streamId);
-      // Only include if job exists AND is still running
-      if (job && job.status === 'running') {
+      // Include running jobs and jobs paused for human review (e.g. tool approval).
+      // A pending-approval job still occupies the user's conversation slot.
+      if (job && (job.status === 'running' || job.status === 'requires_action')) {
         activeIds.push(streamId);
       } else {
         // Self-healing: job completed/deleted but mapping wasn't cleaned - fix it now
diff --git a/packages/api/src/stream/implementations/RedisJobStore.ts b/packages/api/src/stream/implementations/RedisJobStore.ts
index d49affdc28..17f12a20da 100644
--- a/packages/api/src/stream/implementations/RedisJobStore.ts
+++ b/packages/api/src/stream/implementations/RedisJobStore.ts
@@ -215,6 +215,41 @@ export class RedisJobStore implements IJobStore {
       return;
     }
 
+    if (updates.status === 'requires_action') {
+      // Job paused for human review — non-terminal.
+      // Remove from runningJobs so getJobCountByStatus('running') stays accurate,
+      // refresh the hash TTL so the user has the full window to respond, and
+      // leave chunks/runSteps/user-active-set untouched so resume can rebuild state.
+      if (this.isCluster) {
+        await this.redis.srem(KEYS.runningJobs, streamId);
+        await this.redis.expire(key, this.ttl.running);
+      } else {
+        const pipeline = this.redis.pipeline();
+        pipeline.srem(KEYS.runningJobs, streamId);
+        pipeline.expire(key, this.ttl.running);
+        await pipeline.exec();
+      }
+      return;
+    }
+
+    if (updates.status === 'running') {
+      // Resume from requires_action — re-add to runningJobs (idempotent), refresh TTL,
+      // and explicitly clear any stale pendingAction (serializeJob skips `undefined`,
+      // so the only way to remove a hash field is HDEL).
+      if (this.isCluster) {
+        await this.redis.sadd(KEYS.runningJobs, streamId);
+        await this.redis.expire(key, this.ttl.running);
+        await this.redis.hdel(key, 'pendingAction');
+      } else {
+        const pipeline = this.redis.pipeline();
+        pipeline.sadd(KEYS.runningJobs, streamId);
+        pipeline.expire(key, this.ttl.running);
+        pipeline.hdel(key, 'pendingAction');
+        await pipeline.exec();
+      }
+      return;
+    }
+
     if (updates.status && ['complete', 'error', 'aborted'].includes(updates.status)) {
       // Proactively remove from user's job set (requires reading userId from the job hash)
       const job = await this.getJob(streamId);
@@ -428,8 +463,9 @@ export class RedisJobStore implements IJobStore {
 
     for (const streamId of trackedIds) {
       const job = await this.getJob(streamId);
-      // Only include if job exists AND is still running
-      if (job && job.status === 'running') {
+      // Include running jobs and jobs paused for human review (e.g. tool approval).
+      // A pending-approval job still occupies the user's conversation slot.
+      if (job && (job.status === 'running' || job.status === 'requires_action')) {
         activeIds.push(streamId);
       } else {
         // Self-healing: job completed/deleted but mapping wasn't cleaned - mark for removal
@@ -923,6 +959,7 @@ export class RedisJobStore implements IJobStore {
       promptTokens: data.promptTokens ? parseInt(data.promptTokens, 10) : undefined,
       titleEvent: data.titleEvent || undefined,
       replayEvents: data.replayEvents || undefined,
+      pendingAction: data.pendingAction ? JSON.parse(data.pendingAction) : undefined,
     };
   }
 }
diff --git a/packages/api/src/stream/interfaces/IJobStore.ts b/packages/api/src/stream/interfaces/IJobStore.ts
index 3e6786a49b..61fb7a5795 100644
--- a/packages/api/src/stream/interfaces/IJobStore.ts
+++ b/packages/api/src/stream/interfaces/IJobStore.ts
@@ -2,9 +2,13 @@ import type { StandardGraph } from '@librechat/agents';
 import type { Agents } from 'librechat-data-provider';
 
 /**
- * Job status enum
+ * Job status enum.
+ *
+ * `requires_action` is non-terminal: the run has paused for human review
+ * (e.g. tool approval) and is expected to be resumed by an approval route.
+ * Stores must NOT cleanup `requires_action` jobs as if they were complete.
  */
-export type JobStatus = 'running' | 'complete' | 'error' | 'aborted';
+export type JobStatus = 'running' | 'complete' | 'error' | 'aborted' | 'requires_action';
 
 /**
  * Serializable job data - no object references, suitable for Redis/external storage
@@ -53,6 +57,12 @@ export interface SerializableJobData {
   iconURL?: string;
   model?: string;
   promptTokens?: number;
+
+  /**
+   * Set when status is `requires_action`. Describes the human review the
+   * run is waiting on. Cleared by the resume path before the job returns to `running`.
+   */
+  pendingAction?: Agents.PendingAction;
 }
 
 /**
diff --git a/packages/api/src/types/stream.ts b/packages/api/src/types/stream.ts
index dd125a1aab..c93d99d506 100644
--- a/packages/api/src/types/stream.ts
+++ b/packages/api/src/types/stream.ts
@@ -20,9 +20,11 @@ export interface GenerationJobMetadata {
   model?: string;
   /** Prompt token count for abort token spending */
   promptTokens?: number;
+  /** Set when the job is paused for human review (status === 'requires_action') */
+  pendingAction?: Agents.PendingAction;
 }
 
-export type GenerationJobStatus = 'running' | 'complete' | 'error' | 'aborted';
+export type GenerationJobStatus = 'running' | 'complete' | 'error' | 'aborted' | 'requires_action';
 
 export interface GenerationJob {
   streamId: string;
diff --git a/packages/data-provider/src/config.ts b/packages/data-provider/src/config.ts
index 7426d20631..d4c1d7818d 100644
--- a/packages/data-provider/src/config.ts
+++ b/packages/data-provider/src/config.ts
@@ -557,6 +557,40 @@ const remoteApiSchema = z.object({
   auth: remoteApiAuthSchema.optional(),
 });
 
+/**
+ * Decision applied to a tool call before execution.
+ * Mirrors LangChain HITL middleware decision space.
+ *
+ * - `allow`: run the tool without prompting (default).
+ * - `deny`: block the tool; the agent receives a rejection message.
+ * - `ask`: pause the run and require user approval before execution.
+ */
+export const toolApprovalDecisionSchema = z.enum(['allow', 'deny', 'ask']);
+export type ToolApprovalDecision = z.infer<typeof toolApprovalDecisionSchema>;
+
+/**
+ * Per-endpoint tool-approval policy.
+ *
+ * Resolution order for a given tool call:
+ *   1. If the tool name is in `excluded`, the result is `allow`.
+ *   2. If the tool name is in `required`, the result is `ask`.
+ *   3. Otherwise, the result is `default` (which itself defaults to `allow`).
+ *
+ * Tool names are matched as exact strings against the registered tool name.
+ * Pattern matching (`mcp:*`, etc.) is intentionally deferred until we have
+ * a concrete need.
+ */
+export const toolApprovalPolicySchema = z
+  .object({
+    /** Decision applied when no rule matches. Defaults to `'allow'` at policy resolution time. */
+    default: toolApprovalDecisionSchema.optional(),
+    required: z.array(z.string()).optional(),
+    excluded: z.array(z.string()).optional(),
+  })
+  .optional();
+
+export type TToolApprovalPolicy = z.infer<typeof toolApprovalPolicySchema>;
+
 export const agentsEndpointSchema = baseEndpointSchema
   .omit({ baseURL: true })
   .merge(
@@ -579,6 +613,8 @@ export const agentsEndpointSchema = baseEndpointSchema
         })
         .optional(),
       remoteApi: remoteApiSchema.optional(),
+      /** Human-in-the-loop tool approval policy. Off by default. */
+      toolApproval: toolApprovalPolicySchema,
     }),
   )
   .default({
diff --git a/packages/data-provider/src/types/agents.ts b/packages/data-provider/src/types/agents.ts
index 0eab753bd3..7b2c33a7ac 100644
--- a/packages/data-provider/src/types/agents.ts
+++ b/packages/data-provider/src/types/agents.ts
@@ -83,6 +83,16 @@ export namespace Agents {
     auth?: string;
     /** Expiration time */
     expires_at?: number;
+    /**
+     * When set, this tool call is paused for human review.
+     * The presence of this field signals the UI to render approval controls
+     * instead of the in-flight tool execution state.
+     */
+    approval?: {
+      actionId: string;
+      allowed_decisions: ToolApprovalDecision[];
+      description?: string;
+    };
   };
 
   export type ToolEndEvent = {
@@ -262,8 +272,87 @@ export namespace Agents {
     tool_calls?: ToolCallChunk[];
     auth?: string;
     expires_at?: number;
+    /** Approval metadata, set when a tool call is paused for human review. */
+    approval?: {
+      actionId: string;
+      allowed_decisions: ToolApprovalDecision[];
+      description?: string;
+    };
   };
   export type AgentToolCall = FunctionToolCall | ToolCall;
+
+  /**
+   * Human-in-the-loop interrupt category. Currently scoped to tool approvals.
+   * Reserved for future categories like `ask_user_question` (model-driven prompts)
+   * and `respond` (deferred user reply), matching LangChain HITL semantics.
+   */
+  export type HumanInterruptType = 'tool_approval';
+
+  /** Decisions a user can make for a paused tool call. Mirrors LangChain HITL middleware. */
+  export type ToolApprovalDecision = 'approve' | 'reject' | 'edit' | 'respond';
+
+  /**
+   * One pending tool execution awaiting user review.
+   * Field naming mirrors LangChain's `ActionRequest` shape so the same payload
+   * can be forwarded directly when the SDK adopts native HITL primitives.
+   */
+  export interface ToolApprovalRequest {
+    /** Tool name as registered with the agent */
+    name: string;
+    /** Sanitized arguments (no auth tokens / file blobs). May be string or parsed object. */
+    arguments: string | Record<string, unknown>;
+    /** Provider tool_call_id linking this request to the model's tool_use block */
+    tool_call_id: string;
+    /** Optional human-readable description shown alongside the prompt */
+    description?: string;
+  }
+
+  /** Per-tool review configuration: which decisions the user is allowed to make. */
+  export interface ToolReviewConfig {
+    action_name: string;
+    allowed_decisions: ToolApprovalDecision[];
+  }
+
+  /**
+   * Full interrupt payload emitted when an agent run pauses for human review.
+   * Mirrors LangChain's `HumanInterrupt` shape (`action_requests` + `review_configs`).
+   */
+  export interface HumanInterruptPayload {
+    type: HumanInterruptType;
+    action_requests: ToolApprovalRequest[];
+    review_configs: ToolReviewConfig[];
+  }
+
+  /**
+   * Server-side record of a job that is waiting for user input.
+   * Persisted with the job; consumed by approval routes and the status endpoint.
+   */
+  export interface PendingAction {
+    /** Stable identifier used in approval URLs */
+    actionId: string;
+    streamId: string;
+    conversationId?: string;
+    /** Stable per-turn identifier (LangGraph checkpoint_ns) when available */
+    runId?: string;
+    responseMessageId?: string;
+    payload: HumanInterruptPayload;
+    createdAt: number;
+    /** Optional expiry; clients should treat past `expiresAt` as stale */
+    expiresAt?: number;
+  }
+
+  /**
+   * Per-tool decision returned from the approval UI.
+   * `editedArguments` is required when `decision === 'edit'`.
+   * `responseText` is required when `decision === 'respond'`.
+   */
+  export interface ToolApprovalResolution {
+    tool_call_id: string;
+    decision: ToolApprovalDecision;
+    editedArguments?: Record<string, unknown>;
+    responseText?: string;
+  }
+
   export interface ExtendedMessageContent {
     type?: string;
     text?: string;