Choose theme
AI Resources
Manifest
Manifest LLM Gateway connects AI agents and apps to model providers through a shared endpoint. Providers can include API-key services, supported subscriptions and local model servers.
The gateway supplies routing tiers, fallbacks, usage records and request-repair settings. Hosting the gateway yourself and keeping every part of a request local are separate questions. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
One endpoint, configured providers
Point an AI client at the gateway and connect the providers it should use. A harness key identifies that client's configuration.
Why it stands out
Route or name a model
The routing guide separates auto requests, which use tiers and fallbacks, from direct model requests. A matching custom-tier header takes priority over a named model.
Availability
Hosted service or Docker
The project offers a cloud gateway and a self-hosted Docker path. Its repository now lives at mnfst/llm-gateway; the former mnfst/manifest address redirects there.
Why it matters
What makes it useful
To connect an existing agent, distinguish its gateway key from the provider credentials. The self-hosting guide starts with an empty instance: connect a provider, copy the harness's mnfst_ key, then give the client that key and the gateway's base URL. Merely starting the dashboard does not supply a model connection.
What to know
Where it fits
For a task that must name a particular model, the routing guide documents a direct model ID with no routing fallbacks. Its exception matters: a matching custom-tier header still selects that tier's model and fallbacks. For requests using auto, the default tier has a primary model and up to five fallbacks. Check the request's model and tier header together when distinguishing a direct call from a routed one.
Notable points
What stands out
For Gemini thinking-model costs, the installed gateway version matters. The manifest@6.28.1 release notes say earlier Google routes counted candidatesTokenCount without the separate thoughtsTokenCount; the patch includes thinking tokens as output tokens. Those older gateway records therefore missed part of the reported usage. A dashboard cost record is not an independently checked provider invoice.
Before using
What to review
Body recordings and usage metadata are separate. The data guide says the gateway keeps model, token, cost and latency metadata; with logs enabled it also stores full message bodies for each provider attempt. It directs readers who do not want those bodies retained to switch Enable logs off for that harness. That switch does not remove the metadata collection. The request-logs guide also distinguishes each fallback or repaired retry as its own attempt.
Reader fit
Who may find it relevant
Choose self-hosting when the provider is a model server on your own machine: the introduction says the cloud gateway cannot reach it. Inside Docker, the self-hosting guide uses host.docker.internal to reach a server on the host, rather than treating the container's loopback address as that server. It documents a separate hostname for Podman. This is a connection choice; it does not make other configured providers local.
Editorial note
Why LifeHubber lists it
For a self-hosted setup, check hosted Autofix separately from body recording. Its guide says a failed request, including message content, and the provider error go to a hosted healing service for one repair attempt. It documents the harness toggle for stopping request-data transfers, while AUTOFIX_GLOBAL_ENABLED=false stops every healing-service call, including the boot health check. That is an additional service choice even when the gateway itself runs on your machine.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in Ecosystem
Keep browsing this category
Explore more AI ecosystem resources.
Laya
NandhaKishorM/laya
An early Apache-2.0 model family and Python runtime for bounded choice, score, and yes-or-no decisions, with English, multilingual, and task-specialized checkpoints plus a router that selects between them.
CLM
Contrastive-LM/CLM
An Apache-2.0 contrastive model and local server for typed decisions and candidate ranking, with separate state and action embeddings, reusable candidate caches, a browser playground, public reference heads, and fine-tuning tools.
AirLLM
lyogavin/airllm
A Python inference library that streams model weights layer by layer so very large language models can run with far less GPU memory, with PyPI installation, local-model and Hugging Face paths, optional 4-bit or 8-bit compression, CPU support, and an Apple Silicon route.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
