# LiteLLM Cost Tracking — Per-Project LLM Spend Dashboard

> LiteLLM ships a built-in cost dashboard. Track LLM spend by project, user, model, tag. Hard budgets that block at the proxy. SOC2 / SSO via Pro tier.

## Install

Copy the content below into your project:

## Quick Use

1. Have LiteLLM Proxy running with a Postgres database_url
2. Visit `http://localhost:4000/ui` and log in with `UI_USERNAME` / `UI_PASSWORD`
3. Generate a team + user with budgets via the API snippets below

---

## Intro

LiteLLM's cost-tracking layer attributes every LLM call to a project, user, and tag, then surfaces it in a built-in dashboard. Set hard budgets per team — when the budget is hit, the proxy returns 429 instead of forwarding the request. Best for: any team where 'who's burning our LLM budget' is a recurring question. Works with: LiteLLM Proxy (self-host), LiteLLM Cloud (managed). Setup time: 5 minutes (Postgres + Redis + .env).

---

### Enable in proxy config

```yaml
# config.yaml
general_settings:
  master_key: sk-master
  database_url: postgresql://litellm:pass@db:5432/litellm
  store_model_in_db: true
  spend_logs_max_age: "90d"  # auto-prune

litellm_settings:
  callbacks: ["langfuse", "prometheus"]  # optional, ship to your observability
```

```bash
docker compose up -d  # spins up proxy + Postgres
```

### Generate keys with budgets

```bash
# Per-team
curl -X POST http://localhost:4000/team/new \
  -H "Authorization: Bearer sk-master" \
  -d '{"team_alias": "frontend-team", "max_budget": 1000, "budget_duration": "30d"}'

# Per-user (within a team)
curl -X POST http://localhost:4000/user/new \
  -H "Authorization: Bearer sk-master" \
  -d '{"user_id": "alice@acme.com", "team_id": "frontend-team", "max_budget": 50}'

# Generate a key for that user
curl -X POST http://localhost:4000/key/generate \
  -H "Authorization: Bearer sk-master" \
  -d '{"user_id": "alice@acme.com", "max_budget": 50}'
```

When `alice@acme.com` hits $50, all subsequent calls return 429.

### Tag every request

```python
client.chat.completions.create(
    model="claude-3-5-sonnet",
    messages=[...],
    extra_body={
        "tags": ["feature:onboarding", "env:prod", "user-tier:enterprise"],
    },
)
```

The dashboard then shows spend grouped by any tag combination — "how much did onboarding cost last week?" "What's enterprise vs free spend?".

### Built-in dashboard

Visit `http://localhost:4000/ui` (default password from `UI_USERNAME` / `UI_PASSWORD`). Tabs:
- **Spend** — by team, user, model, tag, date
- **Keys** — generate, rotate, revoke
- **Models** — health status, RPM/TPM consumed
- **Logs** — every call with prompt + response (configurable retention)

### Export to your warehouse

```yaml
litellm_settings:
  callbacks: ["s3"]
  s3_callback_params:
    s3_bucket_name: my-llm-logs
    s3_region_name: us-east-1
```

Each call lands as a JSON line in S3 — query with Athena / DuckDB.

---

### FAQ

**Q: Does this require LiteLLM Pro?**
A: No — cost tracking, the dashboard, per-team budgets, and S3 export are all in the open-source proxy. Pro adds SOC2 attestation, SSO/SAML, and managed hosting.

**Q: How accurate is the cost tracking?**
A: LiteLLM uses each provider's official token-count + per-model pricing table to calculate cost per call. For models without published prices (e.g. local Ollama), set `input_cost_per_token` / `output_cost_per_token` in config.yaml or it logs as $0.

**Q: Can I block PII before sending to providers?**
A: Yes — LiteLLM has a Guardrails layer (regex-based or via Presidio / Lakera) that runs before the request goes upstream. Combined with cost tracking, you can also block specific tags from going to specific providers.

---

## Source & Thanks

> Built by [BerriAI](https://github.com/BerriAI). Licensed under MIT.
>
> [BerriAI/litellm — Spend Tracking](https://docs.litellm.ai/docs/proxy/cost_tracking) — ⭐ 17,000+

---

<!-- ZH -->

## 快速使用

1. LiteLLM Proxy 已经跑起来，配好 Postgres database_url
2. 访问 `http://localhost:4000/ui` 用 `UI_USERNAME` / `UI_PASSWORD` 登录
3. 用下面的 API 片段建带预算的团队和用户

---

## 简介

LiteLLM 的成本跟踪层把每次 LLM 调用归属到项目、用户、tag，在自带仪表盘里露出。给每个团队设硬预算 —— 超预算 proxy 直接返 429，不再转发。适合任何「谁在烧我们 LLM 预算」反复出现的团队。需要 LiteLLM Proxy（自托管）或 LiteLLM Cloud（托管）。装机时间 5 分钟（Postgres + Redis + .env）。

---

### 在 proxy 配置里启用

```yaml
# config.yaml
general_settings:
  master_key: sk-master
  database_url: postgresql://litellm:pass@db:5432/litellm
  store_model_in_db: true
  spend_logs_max_age: "90d"  # 自动清理

litellm_settings:
  callbacks: ["langfuse", "prometheus"]  # 可选，发到你的可观测性
```

```bash
docker compose up -d  # 起 proxy + Postgres
```

### 生成带预算的 key

```bash
# 团队级
curl -X POST http://localhost:4000/team/new \
  -H "Authorization: Bearer sk-master" \
  -d '{"team_alias": "frontend-team", "max_budget": 1000, "budget_duration": "30d"}'

# 用户级（团队内）
curl -X POST http://localhost:4000/user/new \
  -H "Authorization: Bearer sk-master" \
  -d '{"user_id": "alice@acme.com", "team_id": "frontend-team", "max_budget": 50}'

# 给该用户生成 key
curl -X POST http://localhost:4000/key/generate \
  -H "Authorization: Bearer sk-master" \
  -d '{"user_id": "alice@acme.com", "max_budget": 50}'
```

`alice@acme.com` 烧到 $50 之后，所有后续调用返 429。

### 给每个请求打 tag

```python
client.chat.completions.create(
    model="claude-3-5-sonnet",
    messages=[...],
    extra_body={
        "tags": ["feature:onboarding", "env:prod", "user-tier:enterprise"],
    },
)
```

仪表盘就能按任意 tag 组合分组看花费 —— "上周 onboarding 花了多少？" "Enterprise 跟 free 花费比？"

### 自带仪表盘

访问 `http://localhost:4000/ui`（默认密码取自 `UI_USERNAME` / `UI_PASSWORD`）。Tab：
- **Spend** —— 按团队 / 用户 / 模型 / tag / 日期
- **Keys** —— 生成、轮换、撤销
- **Models** —— 健康状态、消耗的 RPM/TPM
- **Logs** —— 每次调用带 prompt + 响应（可配置保留期）

### 导出到数仓

```yaml
litellm_settings:
  callbacks: ["s3"]
  s3_callback_params:
    s3_bucket_name: my-llm-logs
    s3_region_name: us-east-1
```

每次调用作为一行 JSON 落 S3 —— 用 Athena / DuckDB 查。

---

### FAQ

**Q: 这功能需要 LiteLLM Pro 吗？**
A: 不需要 —— 成本跟踪、仪表盘、团队级预算、S3 导出都在开源 proxy 里。Pro 加的是 SOC2 认证、SSO/SAML、托管部署。

**Q: 成本跟踪准吗？**
A: LiteLLM 用每个 provider 官方的 token 计数 + 单模型价格表算每次调用成本。没公开价格的模型（比如本地 Ollama），在 config.yaml 里设 `input_cost_per_token` / `output_cost_per_token`，否则记成 $0。

**Q: 能在发到 provider 前拦 PII 吗？**
A: 能 —— LiteLLM 有 Guardrails 层（正则或接 Presidio / Lakera），请求上游前先跑一遍。配合成本跟踪还能阻止特定 tag 去到特定 provider。

---

## 来源与感谢

> Built by [BerriAI](https://github.com/BerriAI). Licensed under MIT.
>
> [BerriAI/litellm — Spend Tracking](https://docs.litellm.ai/docs/proxy/cost_tracking) — ⭐ 17,000+


---
Source: https://tokrepo.com/en/workflows/litellm-cost-tracking-per-project-llm-spend-dashboard
Author: LiteLLM (BerriAI)