Manifest

Evaluates incoming LLM requests across 23 distinct dimensions (token count, reasoning depth, context length, tool requirements, etc.) to compute a complexity score that determines optimal model selection. Routes simple queries to lightweight models and complex reasoning tasks to powerful models via a scoring engine that feeds into the request routing pipeline, reducing inference costs by up to 70% by avoiding overprovisioning.

Acts as an intelligent proxy layer between agents and multiple LLM providers (OpenAI, Anthropic, Ollama, etc.), implementing a proxy pipeline that intercepts requests, applies routing logic, and forwards to the selected provider. If the primary model fails, the system automatically attempts the next model in a pre-configured fallback chain without requiring agent-side retry logic, ensuring resilience across provider outages.

Organizes Manifest as a TypeScript monorepo using npm workspaces and Turborepo for build orchestration, enabling shared type definitions across backend (NestJS), frontend (SolidJS), and plugins. The monorepo structure allows developers to modify shared types and see changes reflected across all packages without separate releases, improving development velocity and type safety.

Uses TypeORM as the ORM layer for PostgreSQL, defining entity schemas for agents, models, subscriptions, analytics, and API keys with TypeScript decorators. Database migrations are version-controlled and run automatically on deployment, enabling schema evolution without manual SQL and supporting rollbacks via migration history.

Implements the Manifest backend using NestJS framework with modular service architecture, organizing code into feature modules (analytics, routing, provider management, OTLP) with dependency injection. Each module encapsulates related business logic (e.g., scoring engine, proxy pipeline, cost tracking) and exposes REST APIs via controllers, enabling clean separation of concerns and testability.

Tracks token usage and inference costs in real-time via an analytics API that aggregates data from all routed requests, stores metrics in PostgreSQL, and enforces hard spending caps per agent or user. When spending approaches or exceeds configured budgets, the system triggers email notifications via a dedicated notification service, preventing runaway costs from unexpected high-volume usage.

Enables agents to leverage existing flat-rate LLM subscriptions (ChatGPT Plus, Claude Pro, GitHub Copilot) by routing requests through provider accounts that have active subscriptions, avoiding per-token billing for models covered by subscriptions. The system maintains a registry of subscription-backed models and prioritizes them in routing decisions when available, effectively converting subscription costs into marginal-cost inference.

Automatically discovers available LLM models from configured providers and synchronizes their pricing data into PostgreSQL via a model discovery pipeline that runs on a scheduled interval. The system maintains a catalog of models with current pricing, context windows, and capabilities, enabling the scoring engine to make cost-aware routing decisions without manual model configuration or stale pricing data.

Ingests observability data via OpenTelemetry protocol (OTLP) from agents and external systems, storing traces and metrics in PostgreSQL, and streams real-time analytics updates to connected dashboards via Server-Sent Events (SSE). This enables live visibility into request routing, token usage, and model performance without polling, implemented as an OTLP service in the NestJS backend.

Provides a web-based dashboard built with SolidJS and Vite that visualizes real-time token usage, costs, model performance, and routing decisions via SSE-streamed analytics data. The dashboard displays interactive charts (via uPlot), cost breakdowns by model/provider, and agent-level metrics, enabling operators to monitor Manifest's behavior and cost optimization effectiveness without CLI tools.

Implements session-based authentication for dashboard users via Better Auth library, and API key-based authentication for agents connecting to Manifest. Agents authenticate by sending API keys in request headers, which are validated against PostgreSQL-stored credentials, enabling secure multi-tenant access where each agent has isolated API keys and cost tracking.

Provides a docker-compose.yml configuration that bundles the NestJS backend, SolidJS frontend, and PostgreSQL 16 database into a single deployable unit, enabling self-hosted deployments without managing infrastructure separately. The Docker image is versioned via the canonical manifest package and supports environment-based configuration for providers, budgets, and SMTP settings.

Provides OpenClaw plugins (manifest and manifest-model-router) that integrate Manifest routing and cost tracking directly into OpenClaw-compatible agents, enabling agents to use Manifest's intelligent routing without modifying core agent code. Plugins expose Manifest's routing and cost APIs as OpenClaw-compatible function calls, allowing agents to invoke routing decisions and track costs as part of their execution flow.