seakee/CPA-Manager-Plus

CPA Manager Plus – What It Is

CPA Manager Plus (CPAMP) is a self‑hosted web panel that sits on top of the CLIProxyAPI (CPA) gateway used to route requests to many large‑language‑model providers (OpenAI, Claude, Gemini, Codex, xAI, Meta/Muse, etc.). It does not forward model traffic itself; instead it observes, stores, and visualises the traffic that CPA already handles.

The panel offers two deployment flavours:

• Lightweight panel – replaces the official CPA UI without adding another service.
• Full‑mode Manager Server – runs in its own container/host, persisting request logs, cost analytics, quota health checks and automation state.

All data stays on your machine (SQLite + encrypted keys); there is no telemetry or required registration.

---

Core Capabilities (as described in the README)

Capability	What You Get
Provider & Credential Management	Add and configure dozens of LLM providers, store OAuth tokens, API keys, model aliases, priorities, and plug‑ins. Supports special flows like Meta’s Device Flow.
Request Monitoring & Failure Diagnosis	Every request that passes through CPA is persisted in a local SQLite DB. You can search by account, API key, status, latency, token count, and view redacted failure evidence. Export/import logs as JSONL.
Cost & Usage Analytics	Break down calls, token usage and estimated cost by model, provider, account, project, channel, time range, etc. Prices are synced from models.dev with fallbacks to LiteLLM and OpenRouter, and you can override them locally.
Account Health & Quota Automation	For Codex and xAI accounts CPAMP reads quota windows, reset evidence and health signals. It can schedule cooldowns, route failed calls to an action queue and help you recover quota automatically.
Production‑Ready Deployment	One‑click Docker image (amd64/arm64) or native packages for Linux, macOS, Windows. Data lives in local files; you can run the manager alongside an existing CPA instance or stand‑alone.
Backup & Security	SQLite files are encrypted with a data.key. Backups require both files. No outbound “phone‑home” traffic; only the provider calls you configure are made.

---

Quick‑Start Overview

1. Install – Run the provided installer script:

   curl -fsSLO https://raw.githubusercontent.com/seakee/CPA-Manager-Plus/main/bin/install-cpamp.sh
   bash install-cpamp.sh
   

   (Add CPAMP_DRY_RUN=1 to preview without installing.)
2. Deploy – Either:
   • Full stack with CPA and CPAMP via Docker‑Compose (see README snippet), or
   • CPAMP only with a single docker run command if CPA is already running.
3. Open the manager UI at http://<host>:18317/management.html, retrieve the admin key from the container logs, and point the UI at your CPA instance.
4. Start adding providers, importing OAuth exports, and watching the dashboards for request health, cost, and quota.

---

Who Might Use This?

• ML Ops teams that run multiple LLM providers behind a single gateway and need a unified view of usage, failures and spend.
• Developers building AI‑augmented products who want to monitor quota health and avoid unexpected cost overruns.
• Enterprises that must keep credential files and request logs on‑premise for compliance.
• Researchers experimenting with many models and wanting quick cost breakdowns per model or project.

---

How It Fits Into the AI Stack

[Your Application] → CPA (CLIProxyAPI) → CPAMP (observability & management) → LLM Provider APIs

CPAMP does not replace the proxy; it simply watches the traffic that CPA already routes, stores it locally, and presents rich dashboards.

---

Documentation & Community

• Live demo (fictional data): https://seakee.github.io/CPA-Manager-Plus/
• Full docs site with guides on panel choice, deployment, provider config, backups, migration from the older CPA‑Manager, and troubleshooting.
• Open‑source MIT license, Docker Hub image seakee/cpa-manager-plus, and a GitHub repo with CI pipelines.

---

TL;DR

CPA Manager Plus is a self‑hosted observability and admin UI for the CLIProxyAPI LLM gateway. It lets you manage credentials, see why requests fail, track token usage and cost, and keep an eye on quota health—all stored locally with encryption and no external telemetry.