Skip to main content

​
What is Manifest?

Manifest is an OpenClaw plugin that brings intelligent LLM routing and comprehensive observability to your AI agents. It intercepts every request, analyzes it through a 23-dimension scoring algorithm in under 2ms, and routes it to the most cost-effective model — cutting costs up to 70%.
Unlike cloud proxies like OpenRouter, Manifest runs locally on your machine. Your prompts and responses never pass through third-party servers, and the routing logic is completely transparent and open source.

​
Why Manifest?

OpenClaw sends all requests to the same model by default, which means you’re summoning expensive flagship models for simple tasks. Manifest solves this by:
  • Smart routing — Analyzing each query and selecting the right model for the job
  • Cost optimization — Automatically choosing cheaper models when quality requirements allow
  • Full transparency — Open-source scoring algorithm with detailed reasoning for every decision

​
The Problem

You ask your agent to summarize a short email. Behind the scenes, it calls GPT-5.3 at $15/million input tokens. The same model processes your complex code generation request an hour later. Both tasks hit the same expensive model, regardless of complexity.

​
The Solution

Manifest analyzes every request across 23 dimensions:
  • Message count — How many turns in the conversation?
  • Text complexity — Technical jargon, code blocks, mathematical notation?
  • Tool usage — Are function calls involved?
  • Momentum — What tier did the last few requests use?
  • Context length — How many tokens are we processing?
Based on this analysis, Manifest routes simple queries to fast, cheap models (DeepSeek-R1, Claude Haiku) and reserves expensive models (GPT-5.3, Claude Opus) for genuinely complex work.

​
Key Features

Smart routing

23-dimension scoring algorithm routes every request to the optimal model in under 2ms, balancing cost and quality automatically.

Real-time tracking

Dashboard shows tokens, costs, and message logs as they happen. See exactly what each model is costing you.

OTLP native

Standard OpenTelemetry ingestion (traces, metrics, logs). Integrates with your existing observability stack.

Privacy by design

Local mode keeps all data on your machine. Cloud mode uses blind proxy architecture — we can’t read your prompts.

300+ models

Supports OpenAI, Anthropic, Google, DeepSeek, xAI, Mistral, Qwen, and more. Pricing database auto-syncs daily.

No coding required

Install as OpenClaw plugin with 3 commands. No SDK integration, no code changes.

​
How It Works

Manifest registers as an OpenClaw plugin and hooks into four key lifecycle events: 
// From packages/openclaw-plugin/src/hooks.ts
message_received     // Start a new trace span
before_agent_start   // Record initial context
tool_result_persist  // Track tool usage
agent_end           // Calculate costs, export telemetry
When your agent makes a request:
1

Interception

The plugin hooks into message_received and creates an OpenTelemetry trace span for the request.
2

Scoring

If the model is set to manifest/auto, the routing engine analyzes the conversation history (last 10 messages) and assigns a tier: fast, balanced, or premium.
3

Selection

Manifest picks the cheapest available model in that tier from your configured providers. If no provider is configured for that tier, it falls back to the next tier up.
4

Telemetry

After the request completes, the plugin calculates token costs using its pricing database (300+ models), attaches metadata to the span, and exports everything via OTLP.

​
Local vs Cloud

Manifest runs in two modes:
Everything stays on your machine:
  • Embedded NestJS server with SQLite database
  • Dashboard at http://127.0.0.1:2099
  • Zero external dependencies
  • Auto-generated API key stored in ~/.openclaw/manifest/config.json
Best for:
  • Privacy-conscious users
  • Offline development
  • Single-machine workflows
openclaw plugins install manifest
openclaw config set plugins.entries.manifest.config.mode local
openclaw gateway restart

​
Privacy Guarantee

In local mode: Your data never leaves your machine. All telemetry, message logs, and cost calculations happen locally. In cloud mode: Only OpenTelemetry metadata is sent:
  • Model name (e.g., gpt-5.3)
  • Token counts (input, output, cached)
  • Latency (response time)
  • Tier (fast, balanced, premium)
What we DO NOT collect:
  • Message content (prompts, responses)
  • User data or PII
  • API keys to LLM providers
The blind proxy architecture means we physically cannot read your prompts — it’s not a trust issue, it’s architectural.

​
Comparison: Manifest vs OpenRouter

FeatureManifestOpenRouter
ArchitectureRuns locally — data stays on your machineCloud proxy — all traffic routes through their servers
CostFree5% fee on every API call
Source codeMIT licensed, fully openProprietary
Data privacy100% local routing and logging (local mode)Your prompts and responses pass through a third party
TransparencyOpen scoring algorithm — see exactly why a model is chosenBlack box routing, no visibility into decisions
Provider support300+ models across 12 providers300+ models via proxy

​
Supported Providers

Manifest tracks pricing for 300+ models across all major LLM providers:
  • OpenAI — GPT-5.3, GPT-4.1, O3, O4-mini + 54 more
  • Anthropic — Claude Opus 4-6, Claude Sonnet 4.5, Claude Haiku 4.5 + 14 more
  • Google Gemini — Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 3 Pro + 19 more
  • DeepSeek — DeepSeek-V3, DeepSeek-R1 + 11 more
  • xAI — Grok-4, Grok-3, Grok-3-mini + 8 more
  • Mistral AI — Mistral Large, Codestral, Devstral + 26 more
  • Qwen (Alibaba) — Qwen3-235b, Qwen3-coder, QwQ-32b + 42 more
  • MiniMax, Kimi, Amazon Nova, Z.ai, OpenRouter, Ollama — All supported
Every provider supports smart routing, real-time cost tracking, and OTLP telemetry.

​
Next Steps

Quickstart

Get Manifest running in under 2 minutes with the quickstart guide.

Installation

Detailed installation instructions for both cloud and local modes.

Routing concepts

Deep dive into how the 23-dimension scoring algorithm works.

Dashboard tour

Explore the real-time analytics, cost tracking, and message logs.

Build docs developers (and LLMs) love

Get started for freeTalk to us