Manifest — Smart LLM Router That Cuts Costs 70%

What it is

Manifest is an intelligent LLM routing layer that sits between your application and LLM providers. It analyzes each incoming request across 23 complexity dimensions and routes it to the cheapest model that can handle it from a pool of 300+ models across 13+ providers including OpenAI, Anthropic, Google, and DeepSeek.

This tool targets teams running production LLM applications who want to reduce API spending without sacrificing output quality. It works as a drop-in proxy that requires minimal code changes.

How it saves time or tokens

Many LLM requests are simple enough for smaller, cheaper models. Manifest identifies these automatically. Simple classification or extraction tasks go to fast models. Complex reasoning tasks go to powerful ones. The routing decision adds under 2ms of latency. Teams report up to 70% cost reduction by avoiding overpowered models for routine requests.

How to use

1. Install Manifest via OpenClaw plugin or run locally with Docker.
2. Open the dashboard at http://127.0.0.1:2099 and add your provider API keys.
3. Point your application to the Manifest proxy endpoint instead of calling providers directly.

# Docker install
docker pull mnfst/manifest
docker run -p 2099:2099 mnfst/manifest

# Or via OpenClaw
openclaw plugins install manifest

Example

Manifest works as a transparent proxy. Your existing API calls route through it:

import openai

# Point to Manifest proxy instead of OpenAI directly
client = openai.OpenAI(
    base_url='http://localhost:2099/v1',
    api_key='your-manifest-key'
)

# Manifest picks the cheapest capable model automatically
response = client.chat.completions.create(
    model='auto',  # let Manifest decide
    messages=[{'role': 'user', 'content': 'Summarize this paragraph...'}]
)

Related on TokRepo

• AI Gateway Providers — Compare LiteLLM and other gateway solutions for multi-provider routing
• AI Tools for API Management — Tools for managing and optimizing API integrations

Common pitfalls

• The 23-dimension scoring is tuned for English text. Non-English requests may get misrouted to weaker models that handle the language poorly.
• Budget controls need careful thresholds. Setting cost caps too aggressively can force all traffic to the cheapest models and degrade quality.
• Automatic fallbacks can mask provider outages. Monitor per-provider success rates separately to catch issues early.