Manifest
Manifest is an open-source, real-time cost observability platform for AI agents and LLM-powered applications. It connects agents and harnesses to any model provider through a gateway layer while capturing telemetry about tokens, costs, messages, and model usage. The project is published as llm-gateway on GitHub under the mnfst organization and is positioned as a way to unify provider access with usage accounting. A core part of the tool is a local-first dashboard that visualizes spend and usage as requests happen, rather than relying on delayed provider billing exports or spreadsheets. It advertises support for 28+ LLM models and accepts telemetry through OTLP ingestion, which is the OpenTelemetry protocol, allowing traces and metrics from instrumented applications to flow into the same observability pipeline. Because it is self-hosted and open source, teams can deploy the gateway and dashboard inside their own infrastructure and keep prompt, message, and cost data on their own systems. This makes it relevant for teams that need per-agent, per-model, or per-session cost visibility without sending operational data to a third-party SaaS vendor. The repository is aimed at developers building agents who want both routing to multiple providers and continuous insight into what those agents cost to run.
- Category
- AI Agents
- Pricing
- Open Source
Problem solved
AI agents and LLM applications commonly call multiple providers and generate costs that are hard to attribute accurately in real time. Provider invoices arrive late and aggregate usage, so teams cannot easily tell which agent, session, or model is driving spend, or catch runaway token usage before the bill grows. Manifest addresses this by pairing a multi-provider gateway with real-time token, message, and cost tracking in a self-hosted dashboard, giving teams immediate visibility into model usage and spend as requests happen.
Use cases
- Monitor real-time token consumption and spend across multiple LLM providers from a single dashboard
- Attribute API costs to specific agents, sessions, or features to find expensive workloads
- Route agent requests to any model provider through one gateway while recording usage telemetry
- Ingest OpenTelemetry (OTLP) traces and metrics from instrumented LLM applications
- Self-host cost and usage observability to keep prompt and message data inside your own infrastructure
- Detect unexpected token spikes or runaway agent loops before provider bills arrive
- Compare cost efficiency across 28+ supported models for the same workload
Audience
- Developers
- ML Engineers
- Data Scientists
- Product Managers
- Startups
Pros
- Open source and self-hosted, so usage and prompt data stay in your own infrastructure
- Real-time dashboard for tokens, costs, messages, and model usage rather than delayed billing reports
- Gateway abstraction connects agents and harnesses to any provider
- Supports OTLP ingestion, fitting existing OpenTelemetry-based observability stacks
- Advertises broad model coverage with 28+ LLM models
Cons
- Cost accuracy depends on provider-reported or locally estimated token counts, which can drift from final invoices
- Requires operating and maintaining a self-hosted gateway plus dashboard
- Limited public detail on licensing, releases, and hosted/managed options
- Value depends on instrumenting agents or routing traffic through the gateway, adding integration work
- Observability features may overlap with existing LLM tracing platforms
Official website: https://github.com/mnfst/llm-gateway
Canonical LambdaBase page: https://www.lambdabase.com/tools/llm-gateway