# PostHog Feature Flags — Gradually Roll Out AI Features

> PostHog Feature Flags toggle features by user, cohort, or percentage. Wrap LLM features behind a flag, A/B different prompts, kill switch on incidents.

## Install

Save in your project root:

## Quick Use

1. PostHog dashboard → Feature Flags → New flag
2. `posthog.get_feature_flag("my-flag", distinct_id=user_id)` (Python) or `posthog.isFeatureEnabled("my-flag")` (JS)
3. Branch your code based on the variant returned

---

## Intro

PostHog Feature Flags let you toggle features by user, cohort, or rollout percentage. Wrap a new LLM feature behind a flag, ramp 1% → 5% → 20% → 100% over a week, kill it instantly if Helicone shows error spikes. Best for: any team shipping LLM features to real users without yolo deploys. Works with: PostHog SDK in 20+ languages, server-side and client-side. Setup time: 5 minutes.

---

### Define a flag in PostHog

PostHog → Feature Flags → New flag:
- Key: `new-summarization-model`
- Rollout: 5% of users (sticky by `distinct_id`)
- Override: enabled for cohort `internal-staff`
- Variant flag (A/B): 50/50 split between `claude-sonnet` and `gpt-4o`

### Server-side (Python)

```python
import posthog

posthog.api_key = os.environ["POSTHOG_API_KEY"]

def summarize(text: str, user_id: str):
    variant = posthog.get_feature_flag(
        "new-summarization-model",
        distinct_id=user_id,
    )

    if variant == "claude-sonnet":
        return claude.summarize(text)
    elif variant == "gpt-4o":
        return openai_summarize(text)
    else:
        return legacy_summarize(text)  # control
```

### Client-side (TypeScript)

```typescript
import posthog from "posthog-js";

posthog.init("phc_...", { api_host: "https://us.posthog.com" });

if (posthog.isFeatureEnabled("show-ai-chat")) {
  renderChatWidget();
}

const variant = posthog.getFeatureFlag("new-summarization-model");
// "claude-sonnet" | "gpt-4o" | undefined (control)
```

### Combine with LLM observability

When you tie flags + LLM traces with the same `distinct_id`, you get end-to-end:

```
PostHog dashboard:
  - Cohort: users on `new-summarization-model = claude-sonnet` (5% rollout)
  - LLM cost p50: $0.012 / call
  - LLM error rate: 0.3%
  - User satisfaction (custom event): 4.6/5

  - Cohort: users on `new-summarization-model = gpt-4o` (5% rollout)
  - LLM cost p50: $0.018 / call
  - LLM error rate: 1.2%
  - User satisfaction: 4.4/5

  → Promote claude-sonnet to 100%
```

### Kill switch

```python
# In your incident playbook
posthog.update_feature_flag("new-summarization-model", { "active": False })
```

Within 30 seconds, all server / client SDKs see the flag disabled, and your code falls back to the control branch. No deploy needed.

---

### FAQ

**Q: How fast is the kill switch?**
A: Server-side SDKs poll every 30 seconds by default. Client-side updates on next page load. For sub-30s reaction, switch to PostHog's `bootstrap` mode or use webhooks to your own kill-switch infrastructure.

**Q: Are flag evaluations free or do they count as events?**
A: Flag evaluations are free. They count toward a separate `decide` quota (much larger than events quota). LLM calls within flag-gated code count as normal LLM events.

**Q: Does this work with Edge / serverless?**
A: Yes — there's a PostHog edge SDK (Cloudflare Workers, Vercel Edge). For very high RPS edge use, prefer client-side flags (lower per-request latency).

---

## Source & Thanks

> Built by [PostHog](https://github.com/PostHog). Licensed under MIT.
>
> [PostHog/posthog](https://github.com/PostHog/posthog) — ⭐ 24,000+

---

<!-- ZH -->

## 快速使用

1. PostHog 仪表盘 → Feature Flags → 新建 flag
2. `posthog.get_feature_flag("my-flag", distinct_id=user_id)`（Python）或 `posthog.isFeatureEnabled("my-flag")`（JS）
3. 按返回的 variant 分支代码

---

## 简介

PostHog Feature Flag 按用户、群组、或推出百分比切换功能。把新 LLM 功能包在 flag 后面，一周内 1% → 5% → 20% → 100% 灰度，Helicone 显示错误飙升时立刻关停。适合任何向真实用户发 LLM 功能、不想 yolo 部署的团队。兼容 20+ 语言的 PostHog SDK，服务端和客户端都行。装机时间 5 分钟。

---

### 在 PostHog 里定义 flag

PostHog → Feature Flags → 新建 flag：
- Key：`new-summarization-model`
- Rollout：5% 用户（按 `distinct_id` 粘性）
- 覆盖：内部员工群组永远启用
- Variant flag（A/B）：50/50 在 `claude-sonnet` 和 `gpt-4o` 间分流

### 服务端（Python）

```python
import posthog

posthog.api_key = os.environ["POSTHOG_API_KEY"]

def summarize(text: str, user_id: str):
    variant = posthog.get_feature_flag(
        "new-summarization-model",
        distinct_id=user_id,
    )

    if variant == "claude-sonnet":
        return claude.summarize(text)
    elif variant == "gpt-4o":
        return openai_summarize(text)
    else:
        return legacy_summarize(text)  # 对照组
```

### 客户端（TypeScript）

```typescript
import posthog from "posthog-js";

posthog.init("phc_...", { api_host: "https://us.posthog.com" });

if (posthog.isFeatureEnabled("show-ai-chat")) {
  renderChatWidget();
}

const variant = posthog.getFeatureFlag("new-summarization-model");
// "claude-sonnet" | "gpt-4o" | undefined（对照组）
```

### 跟 LLM 可观测性结合

flag + LLM trace 用同一个 `distinct_id`，端到端：

```
PostHog 仪表盘：
  - 群组：`new-summarization-model = claude-sonnet`（5% 灰度）的用户
  - LLM 成本 p50：$0.012 / 次
  - LLM 错误率：0.3%
  - 用户满意度（自定义事件）：4.6/5

  - 群组：`new-summarization-model = gpt-4o`（5% 灰度）的用户
  - LLM 成本 p50：$0.018 / 次
  - LLM 错误率：1.2%
  - 用户满意度：4.4/5

  → 把 claude-sonnet 提到 100%
```

### Kill switch

```python
# 事故 playbook 里
posthog.update_feature_flag("new-summarization-model", { "active": False })
```

30 秒内所有服务端 / 客户端 SDK 看到 flag 关停，代码回退到对照分支。不用部署。

---

### FAQ

**Q: Kill switch 多快？**
A: 服务端 SDK 默认每 30 秒轮询。客户端下次页面加载更新。要 30s 以下反应，切到 PostHog 的 `bootstrap` 模式或用 webhook 接到你自己的 kill-switch 基础设施。

**Q: Flag 评估免费还是算事件？**
A: Flag 评估免费。它们算进单独的 `decide` 配额（比事件配额大得多）。flag 门控代码里的 LLM 调用算普通 LLM 事件。

**Q: Edge / serverless 能用吗？**
A: 能 —— 有 PostHog edge SDK（Cloudflare Workers / Vercel Edge）。极高 RPS 的 edge 用法优先客户端 flag（每请求延迟更低）。

---

## 来源与感谢

> Built by [PostHog](https://github.com/PostHog). Licensed under MIT.
>
> [PostHog/posthog](https://github.com/PostHog/posthog) — ⭐ 24,000+


---
Source: https://tokrepo.com/en/workflows/posthog-feature-flags-gradually-roll-out-ai-features
Author: PostHog