AtFlow

See what your LLM calls cost. One command. No signup.

AtFlow is a local observability tool for LLM applications. Point your SDK at it, see your costs, tokens, and latency in real-time.

python -m pip install atflows==0.1b1
atflows

AtFlow 0.1b1 is a beta release. AtFlow uses Bun (>=1.1.0) to run its local server. Install Bun before starting the command. The first launch prepares the bundled runtime in ~/.cache/atflows and needs package network access.

Dashboard: localhost:3000 · Proxy: localhost:8080

---

Quick Start

1. Start AtFlow

The PyPI package and CLI are named atflows; the product is AtFlow.

python -m pip install atflows==0.1b1
atflows

For development from source:

git clone https://github.com/aetna000/atflow.git
cd atflow
bun install
bun run build
bun run dev

2. Point Your SDK

# Python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/v1")

// JavaScript
const client = new OpenAI({ baseURL: 'http://localhost:8080/v1' })

// PHP
$client = OpenAI::factory()->withBaseUri('http://localhost:8080/v1')->make();

3. View Dashboard

Open localhost:3000 to see your traces, costs, and token usage.

---

Who Is This For?

• Solo developers building with OpenAI, Anthropic, etc.
• Hobbyists who want to see what their AI projects cost
• Anyone who doesn't want to pay for or set up a SaaS observability tool

---

Features

Feature	Description
Cost Tracking	Real-time pricing for 2000+ models
Request Logging	See every request/response with latency
Multi-Provider	OpenAI, Anthropic, Gemini, Ollama, Groq, Mistral, and more
OpenTelemetry	Accept OTLP/HTTP traces from LangChain, LlamaIndex, Traceloop, Vercel AI SDK, etc.
Session correlation	Group multi-turn agent runs under one session via session.id (OpenInference / OTel)
Span timeline	Virtualized waterfall view; ~5k spans per trace stays smooth
Zero Config	Just run it, point your SDK, done
Local Storage	SQLite database, no external services

---

Supported Providers

Use path prefixes or the X-AtFlow-Provider header:

Provider	URL
OpenAI	http://localhost:8080/v1 (default)
Anthropic	http://localhost:8080/anthropic/v1
Gemini	http://localhost:8080/gemini/v1
Ollama	http://localhost:8080/ollama/v1
Groq	http://localhost:8080/groq/v1
Mistral	http://localhost:8080/mistral/v1
Azure OpenAI	http://localhost:8080/azure/v1
Cohere	http://localhost:8080/cohere/v1
Together	http://localhost:8080/together/v1
OpenRouter	http://localhost:8080/openrouter/v1
Perplexity	http://localhost:8080/perplexity/v1

---

OpenTelemetry Support

If you're using LangChain, LlamaIndex, or other instrumented frameworks:

# Python - point OTLP exporter to AtFlow
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter

exporter = OTLPSpanExporter(endpoint="http://localhost:3000/v1/traces")

// JavaScript
import { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http'

new OTLPTraceExporter({ url: 'http://localhost:3000/v1/traces' })

Session correlation

If your spans carry one of these attributes, AtFlow groups multiple traces into a single session and exposes them in the Sessions tab:

Convention	Attribute
OpenInference / Phoenix / Arize	session.id (recommended)
LangSmith	langsmith.trace.session_id
Traceloop / OpenLLMetry	traceloop.association.properties.session_id
Vercel AI SDK	ai.telemetry.metadata.sessionId
OTel resource fallback	service.instance.id resource attribute

For chat-thread correlation, set gen_ai.conversation.id (OTel) or traceloop.association.properties.thread_id.

---

Configuration

Variable	Default	Description
PROXY_PORT	8080	Proxy port
DASHBOARD_PORT	3000	Dashboard + OTLP receiver port (npx falls back to 1337)
DATA_DIR	~/.atflows	Data directory
MAX_TRACES	10000	Max traces to retain
VERBOSE	0	Enable verbose logging

Set provider API keys as environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.) if you want the proxy to forward requests.

---

Development

AtFlow is a Bun workspaces monorepo (apps/server, apps/dashboard, plus six packages under packages/). Bun is required.

# Clone and install (one workspace install at root covers every package)
git clone https://github.com/aetna000/atflow.git
cd atflows && bun install

# Server (dashboard on :3000, proxy on :8080)
bun run dev

# Dashboard dev server with HMR (separate terminal, proxies /api + /ws)
bun run dev:dashboard

# Build dashboard for production (outputs to /public/)
bun run build

# Tests
bun run test                # server unit/integration
bun run --filter @atflows/dashboard test    # viewport vitest suite
bun run test:e2e            # Playwright

The dashboard is Svelte 5 + Vite 8 and builds to /public/ at the repo root. The bin entry bin/atflows.js (for the npm workspace) spawns apps/server/src/server.ts directly.

---

Advanced Features

For advanced usage, see the docs/ folder:

• AI CLI Tools - Claude Code, Codex CLI, Gemini CLI
• Observability Backends - Export to Jaeger, Langfuse, Phoenix
• Passthrough Mode - Forward native API formats

---

License

MIT © Helge Sverre and Javad Taghia

Credits

AtFlow is a rebrand and continuation of LLMFlow by Helge Sverre. Original code remains under the MIT license in LICENSE. Rebrand, packaging and dashboard work by Javad Taghia.