Skip to main content

agent-operations

Open-source Agent Operations (AgentOps / LLMOps) toolkit — trace runs, track latency / tokens / cost, and debug tool failures.

Lightweight and self-hostable. Default storage is one SQLite file. No ClickHouse required.

项 值
Python ≥ 3.11
Install pip install agent-operations(或源码 pip install -e .)
Default UI http://127.0.0.1:8787
Default DB ./data/agentops.db

5 分钟上手

1. 安装

python -m pip install agent-operations

从源码开发:

git clone https://github.com/WenSongWang/agent-operations.git
cd agent-operations
python -m pip install -e .
python -m unittest discover -s tests -v

2. 用 Python 写入一条 run

from agentops import connect, ingest_run

conn = connect("./data/agentops.db")
ingest_run(conn, {
    "id": "run-001",
    "project": "demo",
    "agent": "assistant",
    "environment": "dev",          # dev | test | uat | prd
    "name": "answer-question",
    "status": "ok",                # running | ok | error | timeout | cancelled
    "started_at": "2026-10-24T10:00:00+00:00",
    "ended_at": "2026-10-24T10:00:02+00:00",
    "provider": "openai",
    "model": "gpt-4o-mini",
    "input_tokens": 100,
    "output_tokens": 50,
    # cost_usd 可省略:会按内置单价估算
    "steps": [
        {
            "seq": 1,
            "kind": "llm",         # llm | tool | agent | human
            "name": "gpt-4o-mini",
            "status": "ok",
            "started_at": "2026-10-24T10:00:00+00:00",
            "ended_at": "2026-10-24T10:00:02+00:00",
            "input_tokens": 100,
            "output_tokens": 50,
        }
    ],
})
conn.close()

或直接跑示例:

python examples/quickstart.py
python examples/openai_wrap.py

2b. 更省事:用 TracedRun(推荐)

不用手写 running / 终态时间戳;OpenAI 响应可直接抽 token:

from agentops import connect, TracedRun

conn = connect("./data/agentops.db")
with TracedRun(
    conn,
    project="demo",
    agent="assistant",
    environment="dev",
    name="chat",
    task_input="Say hi",
) as tr:
    # resp = client.chat.completions.create(...)
    # tr.record_openai(resp, output_text=resp.choices[0].message.content)
    tr.finish(task_output="hi")
conn.close()

异常退出时会自动记成 status=error。假响应示意:examples/openai_wrap.py。

真实调用(需 OPENAI_API_KEY + pip install openai):

$env:OPENAI_API_KEY = "sk-..."
# 可选:$env:OPENAI_BASE_URL = "https://api.openai.com/v1"
# 可选:$env:OPENAI_MODEL = "gpt-4o-mini"
python examples/openai_live.py

2c. 接 Agentic-Loop

pip install "agent-operations[agentic-loop]"

把 AgentOpsJournal 传给 run_loop(..., journal=...)(与官方 JSONL journal 同接口):

from agentops import connect
from agentops.adapters import AgentOpsJournal

conn = connect("./data/agentops.db")
journal = AgentOpsJournal(
    conn,
    project="demo",
    agent="assistant",
    environment="dev",
    runs_dir="./.agentic-loop/runs",  # 可选:同时写 agentic-loop JSONL
)
# terminal = run_loop(messages, tools=tools, llm=llm, journal=journal)

可运行示意:python examples/agentic_loop_wrap.py。

2d. 接 LangGraph

from agentops import connect
from agentops.adapters.langgraph import traced_invoke

conn = connect("./data/agentops.db")
result = traced_invoke(
    app,  # any object with .invoke / use traced_ainvoke for .ainvoke
    {"messages": [...]},
    conn=conn,
    project="demo",
    agent="support-bot",
    environment="dev",
    name="support",
    config={"configurable": {"thread_id": "1"}},
)

示意(无 LLM):python examples/langgraph_wrap.py。

3. 启动本机服务与页面

agentops-http
# 等价:python -m agentops.http

浏览器打开 http://127.0.0.1:8787

4. 用 HTTP / curl 上报

curl -s http://127.0.0.1:8787/api/runs -Method POST -ContentType "application/json" -Body @'
{
  "id": "run-002",
  "project": "demo",
  "agent": "assistant",
  "environment": "dev",
  "name": "via-http",
  "status": "ok",
  "started_at": "2026-10-24T11:00:00+00:00",
  "ended_at": "2026-10-24T11:00:01+00:00",
  "latency_ms": 1000,
  "input_tokens": 10,
  "output_tokens": 5
}
'@

查询:

curl -s "http://127.0.0.1:8787/api/runs?project=demo"
curl -s "http://127.0.0.1:8787/api/runs/summary?project=demo"
curl -s "http://127.0.0.1:8787/api/failures?project=demo"
curl -s "http://127.0.0.1:8787/api/runs/run-002"

API 一览

方法 路径 说明
GET / 只读调试页
POST /api/runs 写入 / 终态更新一条 run
GET /api/runs 列表(可按 project/agent/environment/status/时间筛选)
GET /api/runs/summary 次数、成功率、token/费用、p50/p95/p99
GET /api/runs/{id} 单条详情(含 steps)
GET /api/failures 工具失败归并

必填字段(run)

id, project, agent, environment, name, status, started_at

终态(ok / error / timeout / cancelled)还要 ended_at。
status=running 时可省略 ended_at;同一 id 再报终态会覆盖更新。

常用可选字段

字段 说明
latency_ms / ttft_ms 耗时;不传则尽量从时间戳 / steps 推算
input_tokens / output_tokens / cached_tokens / reasoning_tokens token;未上报保持空,不是 0
cost_usd 费用;不传则按 provider+model 估算
provider / model 厂商与模型
steps 步骤列表(排错用)
quality_score 0–1
thumbs up / down
prompt_version / template_version 提示词 / 模板版本
tags JSON,如 {"end_user_id":"..."}
extra 备用 JSON

步骤必填:seq, kind, name, status。

可选能力

脱敏

$env:AGENTOPS_REDACT = "hash"   # off | hash | drop

作用于 task_input / task_output、步骤 input / output / traceback、tags.end_user_id。

价目表

未传 cost_usd 时用内置单价估算。自定义 JSON(键为 provider/model):

$env:AGENTOPS_PRICING_PATH = ".\pricing.json"

告警(适合 cron)

$env:AGENTOPS_ALERT_WEBHOOK = "https://example.com/hook"
$env:AGENTOPS_ALERT_FAILURE_RATE = "0.2"
$env:AGENTOPS_ALERT_P95_MS = "3000"
$env:AGENTOPS_ALERT_COST_USD = "10"
agentops-alerts --db .\data\agentops.db

有告警时退出码 2。--dry-run 只打印不 POST。

事件 outbox(Kafka 预留)

$env:AGENTOPS_EVENT_OUTBOX = ".\data\events.jsonl"

每次成功写入追加一行 agentops.run v1 事件(create / finalize)。

API Key(外网绑定前必配)

$env:AGENTOPS_API_KEYS = '{"dev-key":{"projects":["demo"]},"admin":{"projects":["*"]}}'
agentops-http --host 0.0.0.0 --port 8787

请求头:Authorization: Bearer <key> 或 X-API-Key: <key>。
未配置密钥时:仅建议本机 127.0.0.1,接口开放。非 loopback 绑定必须配置密钥。

调试页有 API key 输入框(存浏览器 localStorage),Reload 时自动带 X-API-Key。

限制(请先读)

限制 说明
Alpha 接口可能继续演进;生产使用请自行评估
SQLite 默认单机文件库;多机共享库尚未提供(Postgres 在路线图)
自报数据 延迟 / token / 费用依赖调用方或估算,不是云厂商账单
非全链路 APM 聚焦 agent run + steps,不是通用分布式 tracing
Kafka 目前只有 JSONL outbox,没有内置 broker
页面 只读调试页;可填 API key 后查询
安装渠道 PyPI:pip install agent-operations;版本见 PyPI / tag v0.1.2

更细的后续项见 ROADMAP.md。

开发

python -m pip install -e .
python -m unittest discover -s tests -v

库入口常用符号:connect, ingest_run, list_runs, summarize_runs, get_run, list_tool_failures。

Metadata

Release files for agent-operations 0.1.2

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for agent-operations 0.1.2
File Size Uploaded Provenance
agent_operations-0.1.2.tar.gz 34.1 kB Without provenance Details

Built distribution (wheel)

Table of built distributions (wheels) for agent-operations 0.1.2
File Interpreter ABI Platform
agent_operations-0.1.2-py3-none-any.whl Python 3 none any Details

Total release size: 65.1 kB

Release files / agent_operations-0.1.2.tar.gz

Download URL agent_operations-0.1.2.tar.gz
Size 34.1 kB
Tags Source
SHA-256 checksum
How to use checksums
db466ceeec6ebd4dc6a6cc32a3a8346a875957197be8eabb6f3358b5f08aa705
BLAKE2b-256 checksum
How to use checksums
7cbaa9ce1f7fe34dfd07b071e7e68b14574203da5ce913588cc7a146c3b43843
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.2

Release files / agent_operations-0.1.2-py3-none-any.whl

Download URL agent_operations-0.1.2-py3-none-any.whl
Size 31.0 kB
Tags Python 3
SHA-256 checksum
How to use checksums
766ca63a4dfcabd4a268acd6f737bb382e9bb5d18fbccf9ffe1e23ac0b1bdf2a
BLAKE2b-256 checksum
How to use checksums
1bb763fbefe3928412cc5b3fac5581dd693b9f11c4a701c7835ac11c86a27714
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.2

Security for agent-operations 0.1.2

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page