Local Models Only · No Telemetry · Air-Gap Ready

Private AI Agents on Local Models Only.

TheYgent is a self-hosted, no-code AI orchestration platform built so your data can stay on hardware you control. It ships its own inference engines, keeps every run in your own Postgres, needs no vendor account, and sends no telemetry. This page lists exactly what touches the network, and when — so you can check the claim instead of trusting it.

Local Models, Built In

Most visual AI builders call a model; someone else has to run it. TheYgent runs it. Its inference plane manages three engines — llama.cpp on any machine, MLX on Apple Silicon, and vLLM (experimental) on NVIDIA GPUs — and covers chat, vision, embeddings, speech-to-text, text-to-speech, and image generation. Browse the Hugging Face catalog in the app, see which variants fit your RAM, and install one; the engine starts on the first call and unloads when idle. There is no separate model server to install, expose, or keep running.

Nodes on the canvas reference models by a logical ID such as triage-fast. Nothing forces a hosted model into the picture: an agent built only from local IDs never leaves your machine for inference. If you later register a hosted API, it is one more logical ID, chosen per node — never a default.

Where Your Data Lives

What	Where it stays
Prompts and responses	Browser → control plane → inference plane, over local HTTP. All three listen on `127.0.0.1` by default.
Model weights	Your disk: `~/.theygent/inference/models/`, or wherever you point `THEYGENT_INFERENCE_PLANE_MODEL_DIR`.
Runs, sessions, traces	Your own Postgres. Per-node inputs and outputs are captured per agent policy, with retention and redaction rules you set.
Documents for retrieval	pgvector in your Postgres, embedded by the model you bind — keep the embedding model local and the documents never leave the machine.
Credentials	Two local stores, never inside an agent: hosted-API keys on the inference plane, tool and MCP secrets encrypted in the control plane.
Audio and images	A local artifact directory on the control-plane host — never written into run history or telemetry.

Everything That Touches the Network

The complete list for a running install. Apart from the first row, nothing happens until you open a feature that needs the network, or configure one.

Request	When	What it carries
Public model price list from GitHub	Each time the inference plane starts — its model client library (LiteLLM) fetches it	Nothing of yours. Offline, it falls back to the copy bundled with the library.
Hugging Face	When you browse the model catalog or install a model	Your searches and the files you download. Installed models load from disk afterwards.
MCP registries	When you open the MCP hub browser	Your searches. Point it at a registry you host instead with `THEYGENT_MCP_REGISTRIES`.
Things you configure	When an agent or setting uses them	A hosted API you register, a remote MCP server, an HTTP tool, a site you crawl for retrieval, a sign-in provider, an OpenTelemetry collector, Claude Code traffic you route through the AI Gateway — each sends exactly what you wired it to send.

Not on the list, because TheYgent doesn’t have them: usage analytics, crash reporting, update checks, license activation, or a vendor account. Installation itself — cloning the repo, installing dependencies and engines, pulling container images — needs the network once.

Running Offline or Air-Gapped

Once it is installed, TheYgent runs with no network at all. To set up a host that will never be online:

Install while connected, or on a connected machine you then move across: clone the repo, run make engines and make up, or build the Docker images.
Bring the weights. Install models from the catalog while connected — they register as local files — or copy GGUF files and MLX model folders onto the host and register them by path.
Mirror what you browse. If teams need the MCP hub browser, point THEYGENT_MCP_REGISTRIES at a registry you host; it doubles as an allowlist.
Run durable work in-house. On a server, run the standalone durable worker next to the control plane; runs, approvals, and schedules then survive restarts without any outside service.

The startup price-list request simply fails and falls back to the bundled copy; nothing else expects to be online.

Honest About the Promise

TheYgent doesn’t stop you from sending data out — you can register a hosted API, call a web service, or export traces whenever you choose. The promise is narrower and checkable: you own every hop. TheYgent is never a middleman you didn’t choose, and nothing crosses the wire unless you pointed it there.

The control plane requires sign-in — a fresh install’s API is closed until the first-run wizard creates a local admin. The inference plane has no sign-in of its own and listens on 127.0.0.1; keep port 8081 off untrusted networks.

Questions

Asked and Answered.

Is there a no-code AI agent builder that only uses local models?

Yes. TheYgent is a self-hosted, no-code AI agent builder with built-in local inference: llama.cpp on any machine, MLX on Apple Silicon, and vLLM (experimental) on NVIDIA GPUs. You build agents on a visual canvas and install models from an in-app catalog. Hosted APIs are optional and off until you register one.

Can TheYgent run completely offline?

Yes, once it is installed. Install TheYgent and your models while connected — or copy the weights onto the host and register them by path — and everything runs without network access: the canvas, the engines, Postgres, durable runs, and retrieval with a local embedding model. The only automatic request, a public price list fetched at startup, falls back to a bundled copy when offline.

Does TheYgent collect telemetry?

No. TheYgent has no usage analytics, crash reporting, update checks, or license activation, and it needs no vendor account. Traces stay in your own Postgres; exporting them over OpenTelemetry is off until you set OTEL_EXPORTER_OTLP_ENDPOINT to a collector you run.

Where does TheYgent store my data?

On machines you control. Agents, runs, sessions, traces, and retrieval documents live in your own Postgres; model weights and the model registry live on the inference plane’s disk; audio and image artifacts live in a local directory. Credentials are kept in two local stores and never written into an agent.

Can I still use a hosted model when I need one?

Yes. Register any OpenAI-compatible API under a logical ID and use it on the nodes you choose; one graph can mix local and hosted models. Calls to that model go straight from your machine to the provider you picked, never through a TheYgent server.

Keep Reading

mlx