See what your LLM calls cost. One command. No cloud signup.
AtFlows 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
atflows initAtFlows 0.1.3 adds continuity observability, grouped event charts, and AtMem-aligned typography. It retains optional AtMem-owned dashboard login. AtFlows uses Bun (>=1.1.0); install it before standalone startup, or use AtMem's managed setup. The first launch prepares the bundled runtime in ~/.cache/atflows and needs package network access.
atflows init creates a temporary Local Administrator password, starts the server, and opens the sign-in page with it filled in. Choose a permanent password to finish setup. If access is lost, run atflows users recover-administrator; it issues a new temporary password. OTLP ingestion and the model proxy remain available to configured clients.
Run atflow or atflow status to list live dashboard and proxy addresses. atflows status does the same; atflows starts a server as before. Use atflow start to start one through the short command.
Dashboard: localhost:1337 by default · Proxy: localhost:8080
Both listeners bind to 127.0.0.1 by default. DASHBOARD_HOST and PROXY_HOST can change their bind addresses for a trusted deployment; OTLP ingestion has no built-in authentication. To use one AtMem account for both dashboards, start the AtMem dashboard, set ATFLOWS_ATMEM_AUTH_URL to its numeric loopback URL, and run atflows init. Keep the setting for every AtFlows start; the mode is read at startup. Open both dashboards using 127.0.0.1. AtMem then owns users, roles and passwords. AtFlows' local accounts are inactive until you restart without the setting. See users and access.
See the setup and connection guide and the integration catalog for provider routes, telemetry, and tool-specific instructions. Working recipes cover Claude Code, OpenClaw, LangChain, Pydantic AI, and AtBots. Choose a connection in the dashboard for your running addresses and guided steps. See the 0.1.3 release notes for changes and current limitations.
AtFlows 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/atflows.git
cd atflows && bun install
# Server (dashboard on :1337, 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 # PlaywrightThe 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.
For advanced usage, see the docs/ folder:
- Observability Backends - Export to Jaeger, Langfuse, Phoenix
- Architecture pathway - Production storage, CLI/MCP, AtMem/AtBot, and Jev requirements
- Spec Kit workflow - Constitution and feature specifications
AtFlows changes by Javad Taghia are licensed under Apache 2.0. The inherited LLMFlow code remains subject to its original MIT license and Helge Sverre copyright notice. Both license texts and the NOTICE are included with the Python package.
AtFlows is a rebrand and continuation of LLMFlow by Helge Sverre. Original code remains under the MIT license in LICENSE-MIT-LLMFLOW. Rebrand, packaging and dashboard work by Javad Taghia.