AIHelms

Enterprise AI Resource Management Platform

Built on top of LiteLLM · Model distribution, AI identity, cost control, and governance

[![Release](https://img.shields.io/badge/Release-v0.1.25-brightgreen.svg)](https://github.com/beizhu-1209/AIHelms/releases/tag/0.1.25) [![Docker](https://img.shields.io/badge/Docker-Compose-2496ED.svg)](https://www.docker.com/) [![Python](https://img.shields.io/badge/Python-3.11+-3776AB.svg)](https://www.python.org/) [![FastAPI](https://img.shields.io/badge/FastAPI-009688.svg)](https://fastapi.tiangolo.com/) [![Vue](https://img.shields.io/badge/Vue-3.4+-4FC08D.svg)](https://vuejs.org/) [![TypeScript](https://img.shields.io/badge/TypeScript-3178C6.svg)](https://www.typescriptlang.org/) [![TailwindCSS](https://img.shields.io/badge/TailwindCSS-06B6D4.svg)](https://tailwindcss.com/) [![PostgreSQL](https://img.shields.io/badge/PostgreSQL-16+-336791.svg)](https://www.postgresql.org/) [![Redis](https://img.shields.io/badge/Redis-7+-DC382D.svg)](https://redis.io/) [![vLLM](https://img.shields.io/badge/vLLM-FF6F00.svg)](https://docs.vllm.ai/) [![SGLang](https://img.shields.io/badge/SGLang-8B5CF6.svg)](https://github.com/sgl-project/sglang)

| English | 简体中文 |

[What Is It](#what-is-it) · [Why AIHelms](#why-aihelms) · [Quick Start](#quick-start) · [Deployment](#deployment) · [Cookbook](http://www.aihelms.cn/docs)

--- ## What Is It AIHelms is an AI resource management platform for enterprises. It helps companies **manage AI assets, control AI spend, and measure the value of AI**. - **For leadership** — the platform records usage, calculates cost, and generates reports, so the return on your AI investment is visible. - **For administrators** — onboard models, assign identities, and set budgets from a single admin console. - **For employees** — get an AI identity from the user portal and use it with any client. ![Architecture](docs/struct.png) ## Why AIHelms ### Cost you can see and control The biggest worry with enterprise AI is spending money without knowing where it goes. AIHelms builds a full cost model around that: - **Dual pricing, internal and external** — external cost (what the provider charges) and internal settlement price (what you bill departments) are tracked separately. Pass cost through as-is, or add a markup for internal settlement. - **Multi-dimensional budgets** — set budgets per person, department, project, or model, with soft limits (alerts) and hard limits (cut off on overspend). - **Real-time attribution** — every call traces back to a specific person, department, project, and model, down to the token level. ### One identity, one entry point - One AI identity (API key) per employee, one entry point to every AI resource. - AI employees (in progress). ### AI value, quantified - **Coverage** — department coverage, activation rate, model usage distribution. - **Activity** — daily call volume, tokens per person, usage trends. - **Cost efficiency** — department cost ranking, model cost-effectiveness, budget burn, internal vs. external cost gaps. - **Efficiency reports** — generated weekly and monthly, with drill-down by department, project, and model. ### Security and audit - Sensitive data detection. - High-risk prompt interception. - Admin action tracking. ### Centralized resource management - **Model management** — onboard multiple providers behind one gateway, load-balance across deployments of the same model, support both OpenAI and Anthropic formats. - **Skill & MCP** — register, approve, and distribute internal AI tools in one place. - **Fine-grained permissions** — who can use which model, how much quota, and when it expires, all configurable. ## Changelog Recent versions are summarized from [GitHub Releases](https://github.com/beizhu-1209/AIHelms/releases). See the release page for the full changelog. | Version | Date | Summary | |------|----------|----------| | `0.1.26` | 2026-10-26 | Added `LOG_DETAIL_CONTENT_VISIBLE_TO_ADMIN` to control whether regular admins can view request, response and metadata in LLM/MCP log details; super admins can always see them. | | `0.1.25` | 2026-09-26 | Fixed saving Anthropic-format model deployments hanging and returning 504. | | `0.1.24` | 2026-09-10 | Fixed model publishing authorization: grants now follow the selected department scope, unpublishing revokes access and invalidates approvals, and new users get keys scoped to their visibility. | | `0.1.23` | 2026-09-04 | update Hub OpenAPI , plz use V2 endpoint| | `0.1.22` | 2026-08-25 | Hub now supports email sign-in; export task history auto-cleanup is now configurable; reworked the MCP log sync window to reuse the LLM config in `.env`. | | `0.1.21` | 2026-08-13 | Fixed user management role/status filter pagination issue. Update export | | `0.1.20` | 2026-08-04 | Fixed log sync issues; fixed timestamp display; updated Model Square access info. | | `0.1.19` | 2026-07-23 | Added Hunyuan and Grok provider support; reworked static asset (icon) hosting; enhanced batch operations for some APIs; upgraded the LiteLLM model-management base to v1.93. | | `0.1.18` | 2026-07-22 | Added a Token column to cost detail by department/project with inline drill-down into members; added Token usage stats and a Top 10 people leaderboard to Efficiency overview, cost, budget and the dashboard; reworked the budget detail "Key budget" tab into "Per-person budget". | | `0.1.17` | 2026-07-17 | Unified department and project filters across AI Efficiency and fixed cost details; added fuzzy search to log and admin filters with correct person-to-Key linkage; added Workbuddy to the user-portal client guidance. | | `0.1.16` | 2026-07-14 | Added enterprise license management and platform branding customization. | | `0.1.15` | 2026-07-013 | Fix bug . | | `0.1.14` | 2026-07-07 | AI Policies adds LLM-based review. | | `0.1.13` | 2026-07-03 | New AI Policies security review: scan a Skill in one click, flag malicious instructions, data exfiltration, and excessive permissions, and generate a report. | | `0.1.12` | 2026-07-01 | Resource requests support batch approval; Skill / MCP add author info, with usage shown in the user portal. | | `0.1.11` | 2026-06-29 | Fixed log mismatches for long agent logs such as codex /goal. | | `0.1.9` | 2026-06-27 | Added Xiaomi MiMo model onboarding and model logo selection; more complete AI identity rate limiting (total plus per-model). | | `0.1.8` | 2026-06-23 | Support for LiteLLM Anthropic-format access through vLLM. | | `0.1.7` | 2026-06-23 | Fixed several cost calculation and log statistics issues for more accurate accounting. | | `0.1.6` | 2026-06-16 | Improved model connectivity testing. | | `0.1.5` | 2026-06-10 | Upgraded the admin dashboard and the cost, efficiency, and budget analysis experience. | | `0.1.4` | 2026-06-04 | Fixed several issues and merged community contributions. | ## Features ### Admin console For IT administrators and AI owners, covering the full resource management workflow. | | | |:---:|:---:| | **Dashboard** — today's calls, cost, and to-dos at a glance | **Model management** — provider onboarding, multiple deployments, connectivity testing | | | | | **AI identity** — key issuance, budget allocation, lifecycle | **AI efficiency** — cost analysis, coverage, efficiency reports |
Admin console feature list | Module | Capabilities | |------|------| | Model management | Provider management, credential management, model registration, multi-deployment load balancing, routing strategy, connectivity testing | | AI identity | Department / project / people management, key issuance, budget quota, rate limiting, model allow-list | | AI marketplace | Skill publishing, MCP Server registration and discovery, resource request approval | | AI efficiency | Cost analysis (internal / external), adoption coverage, budget control, efficiency reports | | Security | Admin audit log, action tracking | | Agent center | Agent creation, configuration, lifecycle management |
### User portal (AI Hub) For all employees. Simple and direct: get an AI identity and start using it. ![User portal](docs/web.png) - View your AI identity and copy the key and endpoint in one click. - Browse the model square and pick a model ID by scenario. - Browse the AI marketplace and request access to Skills / MCP. - Check this month's usage and budget burn. ## Quick Start ### Requirements - Docker 20.10+ & Docker Compose v2+ - 4 cores, 8 GB RAM (minimum) ### Deployment ```bash git clone https://github.com/beizhu-1209/AIHelms.git cd AIHelms cp .env.example .env # Edit .env and fill in your secrets docker compose up -d ``` ### Access | URL | Description | |------|------| | `http://your-host` | User portal | | `http://your-host/admin` | Admin console | | `http://your-host/api/docs` | API docs | Default admin: `admin` / password is `SUPER_ADMIN_PASSWORD` in `.env`. ### Three steps to get started ``` 1. Admin console → Providers → Add credential 2. Admin console → Models → Create model → Link credential → Publish 3. User portal → My AI Identity → Copy key and endpoint → Configure your client ``` ## Tech Stack | Layer | Technology | |------|------| | Backend | Python 3.11+, FastAPI, Gunicorn, Celery | | Frontend | Vue 3.4+, TypeScript, Vite 5+, TailwindCSS | | Database | PostgreSQL 16+ | | Cache | Redis 7+ | | Deployment | Docker Compose, Nginx |
Environment variables | Variable | Description | |------|------| | `POSTGRES_PASSWORD` | Database password | | `SECRET_KEY` | JWT signing key | | `SUPER_ADMIN_PASSWORD` | Super admin initial password | | `WEB_PORT` | Web external port | See `.env.example` for the full list.
Development ```bash ./dev/setup # First-time setup ./dev/start-docker-compose # Start middleware ./dev/start-api # Start backend (hot reload) ./dev/start-web # Start frontend (HMR) ``` See [Development and release workflow](docs/INTERNAL.md).
## Roadmap ### AI Lab | Feature | Summary | |------|------------| | Reports | Aggregate AI efficiency, cost, and budget analysis into a management-facing report view. | | AI health | Monitor MCP upstreams, model deployments, the Docker environment, and efficiency data refresh status in one place. | | Sensitive data detection | Detect sensitive information in AI interactions to support alerts, audit, and handling. | | A2A | Plan collaboration, call relationships, and task chaining between agents. | | Context caching | Cache frequent context and session material to reduce repeated input and call cost. | | Policy management | High-risk detection, including geo-fencing and prompt injection. | | File processing | Parse, chunk, index, and preprocess uploaded files for downstream AI capabilities. | ## Contributing 1. Fork this repository. 2. Create a branch: `git checkout -b feature/xxx`. 3. Commit to the dev branch: `git commit -m "feat: describe your feature"`. 4. Push and open a Pull Request. See [Development and release workflow](docs/INTERNAL.md) for detailed conventions. ## Community - WeChat group: scan to join. WeChat group - Business inquiries: [email us](mailto:jiangshiwei@microbaton.com) - Issues: [GitHub Issues](https://github.com/beizhu-1209/AIHelms/issues) ## License AIHelms is licensed under [GPL-3.0](LICENSE), and runs on the LiteLLM community edition (MIT License). If your organization's policies do not allow GPL-3.0 licensed software, or you want to use AIHelms without the obligations of GPL-3.0, email us at [Mail](mailto:jiangshiwei@microbaton.com)for a commercial license.