# Manifest

> Manifest is an open-source LLM router that cuts AI agent costs by up to 70%. It is OpenAI-compatible and works with personal AI agents, coding assistants, and app AI SDKs.

Manifest analyzes each query locally in under 2ms and routes it to the right model that can handle it. The routing engine runs entirely on your machine. The cloud version uses a blind proxy that redirects requests to the right provider without reading them. Your data stays yours by architecture, not by policy.

Manifest supports 500+ models across every major provider. Budgets and rate limits, custom providers, model fallbacks, and spend tracking are all built in.

## API contract

- Cloud base URL: `https://app.manifest.build/v1`
- Self-hosted base URL: `http://localhost:2099/v1`
- Auth header: `Authorization: Bearer mnfst_<key>`
- Get a key: https://app.manifest.build
- Request model: `manifest/auto` (the only accepted value; the chosen provider model is returned in the response, e.g. `auto→claude-sonnet-4`. Provider-specific names like `gpt-4o-mini` are not accepted as request models.)
- Endpoints:
  - `POST /v1/chat/completions` (OpenAI-compatible)
  - `POST /v1/messages` (Anthropic-compatible)
- Client SDK: none needed — use the official `openai` or `@anthropic-ai/sdk` packages with `baseURL` set.

## How it works

Manifest sits between an AI agent and the LLM provider as an OpenAI-compatible endpoint. For every request:

1. The router reads the prompt locally and scores it across 23 dimensions (length, complexity, code/math/reasoning intent, required context window, expected output length, latency tolerance, and more).
2. It picks the right model that meets the request's requirements from the configured pool.
3. It forwards the request to that provider using the existing OpenAI-compatible API surface.
4. It records the cost, tokens, latency, and outcome for the dashboard.

The local routing decision adds under 2ms. No prompt content leaves the machine in local mode. In Manifest Cloud, the proxy never reads request bodies — it only routes by metadata.

## What's included

- **Local LLM router** — Docker image, OpenAI-compatible at `http://localhost:2099/v1/chat/completions`. Drop-in replacement for `OPENAI_BASE_URL`.
- **500+ models** across OpenAI, Anthropic, Google, Mistral, DeepSeek, Alibaba, Moonshot, MiniMax, xAI, Z.ai, Groq, Together, Fireworks, OpenRouter, and self-hosted endpoints.
- **Provider tokens** — bring your own keys, or use Manifest Cloud's pooled access.
- **Budget guardrails** — daily/weekly/monthly caps, per-key limits, hard stops with email alerts.
- **Rate limiting** — token-bucket per provider/model/key, automatic fallback on 429.
- **Model fallback chains** — declarative fallback order on errors or quota exhaustion.
- **Spend dashboard** — real-time cost, tokens, requests, errors, and per-model breakdowns.

## Compatible agents and SDKs

Personal AI agents:
- OpenClaw
- Hermes Agent
- Nanobot
- Craft Agent

Coding assistants:
- Claude Code
- OpenCode

App AI SDKs:
- OpenAI SDK (Python, Node, Go, Rust, etc.)
- Anthropic SDK
- Vercel AI SDK
- LangChain
- LlamaIndex
- LiteLLM
- Anything that speaks the OpenAI-compatible chat completions API

## Quickstart (local)

```bash
bash <(curl -sSL https://raw.githubusercontent.com/mnfst/manifest/main/docker/install.sh)
# Open http://localhost:2099 and sign up — the first account becomes the admin.
# Use http://localhost:2099/v1 as your OpenAI base URL.
```

Set `OPENAI_API_KEY` to your Manifest agent key (starts with `mnfst_`). Manifest uses your configured provider tokens for upstream calls.

## Quickstart (cloud)

1. Sign up at https://app.manifest.build/
2. Add your provider keys (or use the included pool)
3. Copy your Manifest API key (starts with `mnfst_`)
4. Point your agent at `https://app.manifest.build/v1` with that key

The blind-proxy architecture means Manifest Cloud forwards your requests to the chosen provider without reading the body. Routing decisions happen on metadata only.

## Architecture (cloud)

```
[ Your agent ]
      |
      v
[ Manifest API (HTTPS) ]
   - auth + budget check (metadata only)
   - routing decision (metadata only)
   - blind proxy passes body to upstream
      |
      v
[ Provider: OpenAI / Anthropic / ... ]
```

Request bodies are never logged. Aggregate cost/latency telemetry is recorded per request.

## Pricing

- **Local** — free, MIT licensed, no telemetry.
- **Cloud** — usage-based; routing is free, you pay the upstream provider cost (passed through) plus a small platform fee. See https://manifest.build/ for current pricing.

## Docs

- [Documentation](https://manifest.build/docs/introduction/): Full documentation for installing, configuring, and using Manifest
- [Set Up Manifest Cloud](https://manifest.build/blog/setup-manifest-cloud/): Set up Manifest Cloud and start routing agent requests
- [What is an LLM Router?](https://manifest.build/blog/what-is-an-llm-router/): Rule-based vs. AI-powered routing and what it means for autonomous agents
- [Free Models](https://manifest.build/free-models/): LLM APIs with permanent free tiers
- [Providers](https://manifest.build/providers/): Every provider Manifest supports, with live token counts and active models
- [Agents & SDKs](https://manifest.build/agents/): Setup guides for OpenClaw, Hermes, Claude Code, OpenCode, and every supported SDK
- [Solutions](https://manifest.build/solutions/ai-builders/): Manifest for AI builders, teams, and platforms
- [LLM Error Catalog](https://manifest.build/errors/): Provider errors with live frequency data, causes, and fixes
- [Pricing](https://manifest.build/pricing/): Local is free forever; Cloud and Enterprise plans
- [Self-hosted](https://manifest.build/self-hosted/): Run the router on your own infrastructure with Docker
- [Changelog](https://manifest.build/changelog/): New features, fixes, and improvements, released on GitHub

## Product

- [Manifest Cloud](https://app.manifest.build/): Sign up and start routing requests
- [GitHub Repository](https://github.com/mnfst/manifest): Open source codebase, issues, and contributions
- [npm: @manifestbuild/manifest](https://www.npmjs.com/package/@manifestbuild/manifest): Local CLI

## Community

- [GitHub Discussions](https://github.com/mnfst/manifest/discussions): Questions, ideas, and community threads
- [Discord](https://discord.gg/FepAked3W7): Real-time chat and support
- [Blog](https://manifest.build/blog/): Weekly updates and long-form posts
- [RSS](https://manifest.build/blog/rss.xml): Blog feed

## Company

- [About](https://manifest.build/about/): The team behind Manifest
- [Privacy Policy](https://manifest.build/privacy/): How we handle your data
- [Terms of Service](https://manifest.build/terms/): Terms and conditions

## Selected blog posts

The full, always-current list of posts lives in https://manifest.build/llms.txt and at https://manifest.build/blog/.

### The Ultimate Guide to AI Subscription Plans
Connect Claude Max, ChatGPT, Copilot, Gemini, Grok and more to your AI agent through Manifest, and use the plans you already pay for instead of per-request API fees. https://manifest.build/blog/subscription-plans-guide/

### Everyone is building LLM routers, we deprecated ours
Why Manifest shut down its complexity-based LLM router: prompt complexity can't be inferred from the prompt, caching beats routing on cost, and routing breaks behavior consistency. https://manifest.build/blog/why-we-deprecated-our-llm-router/

### The errors that actually break LLM agents in production
The six plumbing failures seen most in Manifest's logs, and how to handle them. https://manifest.build/blog/llm-agent-errors-production/

### Run Claude Code on your ChatGPT Plus subscription
Route Claude Code requests through an existing ChatGPT Plus subscription instead of paying for API keys on top. https://manifest.build/blog/run-claude-code-on-chatgpt-plus/

### 10 Ways To Reduce Your LLM API Costs
Practical ways to cut LLM inference costs without hurting output quality. https://manifest.build/blog/reduce-ai-inference-costs/

### Don't Hardwire Your Agents to One Model
Fable 5 doubled prices then access was suspended in three days — the case for routing over single-model dependency. https://manifest.build/blog/fable-5-routing/

### This will get you banned from your ChatGPT subscription
Subscriptions are cheap inference but strictly personal; what usage patterns cross the line. https://manifest.build/blog/banned-from-chatgpt-subscriptions/

### Anthropic kicked autonomous agent users off their subscription
Claude Max subscriptions stopped covering agent usage. Walks through alternatives: bring-your-own keys, free-tier providers, intelligent routing through Manifest. https://manifest.build/blog/anthropic-kicked-openclaw-users-off-subscription/

### How to Stop Burning Money on AI Agents
Cost-saving strategies that actually work, drawn from conversations with 100+ agent users. Covers caching, prompt compression, model selection, and routing. https://manifest.build/blog/how-to-stop-burning-money-on-openclaw/

### Introducing Manifest
Launch announcement. Why agent costs are unpredictable, how routing helps, and where Manifest fits in the stack. https://manifest.build/blog/introducing-manifest/

### What is an LLM Router?
Rule-based vs. AI-powered routing and what it means for autonomous agents. The case for local-first routing. https://manifest.build/blog/what-is-an-llm-router/

### Set Up Manifest Cloud
Step-by-step setup for Manifest Cloud, with screenshots. https://manifest.build/blog/setup-manifest-cloud/

### Awesome Free LLM APIs
All the LLM APIs with permanent free tiers, updated regularly. https://manifest.build/blog/awesome-free-llm-apis/

### The Claw Market Map, Q1 2026
Key players across managed hosting, LLM routing, security, developer tools, and more. https://manifest.build/blog/claw-market-map-q1-2026/

### Hyper Specialization: Stockfish, Adam Smith and the AI Era
Why hyper specialization, not AI orchestration, may save our jobs. https://manifest.build/blog/hyper-specialization-stockfish-adam-smith/

### My AI agent wouldn't stay autonomous
Skill files, cron jobs, memory protection, and anti-drift rules for keeping agents on task. https://manifest.build/blog/my-openclaw-agent-wouldnt-stay-autonomous/

### How To Shut Up OpenClaw CLI Banner
Disable the CLI banner jokes in one command. https://manifest.build/blog/shut-up-openclaw/

## License

MIT — see https://github.com/mnfst/manifest/blob/main/LICENSE