Everything you need to know about Mealissa — what it is, how it's secured, and how to get your local LLM online.
Mealissa is a secure, OpenAI- and Anthropic-compatible gateway that gives self-hosted local LLMs a globally accessible API endpoint, with prompts encrypted in transit and never stored. Run any Ollama model on your own machine and reach it from anywhere through one stable endpoint.
Local LLMs are isolated by default — machines behind home or office routers have no safe, standard way to be reached from outside. Mealissa bridges your machine and the internet: your model becomes a production-grade API endpoint without open ports, certificates to manage, or your data resting in someone else's cloud.
Developers and organizations running self-hosted models who want to reach them from apps, scripts, other machines, or teammates — without cloud GPU costs or vendor lock-in.
Exposing Ollama directly means an open, unauthenticated port on the internet. Tunnels and VPNs like ngrok or Tailscale solve transport but nothing else. Mealissa adds what those don't: Anthropic- and OpenAI-compatible APIs so existing SDKs work unchanged, license-key authentication on every call, a usage dashboard, multi-node groups — and zero prompt storage.
Today Mealissa fronts Ollama specifically, the agent runs on Windows 10/11 and Linux (x64), each plan caps the maximum model parameter size, and routing through the gateway adds some latency compared to a call on your own network.
Your prompt and response content is never written to disk or logs — requests are processed in memory only long enough to route them. What is kept is usage metadata (model name, token counts, latency, status) for billing and your dashboard, as described in the privacy policy.
All traffic is encrypted in transit with TLS 1.2/1.3. The agent connects outbound over an encrypted WebSocket (wss://) and refuses to start with an unencrypted URL; API calls are HTTPS and authenticated with your license key.
Prompts pass through the gateway encrypted in transit and are never stored or logged. Only usage metadata — never conversation content — is recorded.
You do. Models run on your hardware with your weights; Mealissa handles only the secure exposure.
You create and manage license keys in the cloud console. Every API call authenticates with a key sent as a bearer token; keys can be listed and validated from the console and are revoked when an account closes.
Yes — every request must present a valid license key, and unauthenticated calls are rejected. Finer-grained controls such as IP allowlists are not currently offered.
The mealissa-llm-agent runs next to Ollama on your machine and opens an outbound, encrypted WebSocket to the Mealissa cloud, registering your node and model. Callers send standard Anthropic or OpenAI API requests to your endpoint; the gateway forwards each request over that persistent connection to your agent, which queries Ollama and returns the answer. Nothing is persisted along the way. See how the protocol layer works.
No. The agent's connection is outbound-only — no port forwarding, no static IP, no router configuration.
One encrypted round trip through the gateway on top of your model's inference time; the exact overhead depends on your distance from the relay. The usage dashboard shows per-request latency so you can measure it for your own setup.
Not currently — the mealissa-llm-agent is distributed as a free download for Windows and Linux.
Updates are manual today: download the new version from the downloads page, stop the agent, replace it, and start it again. Your node re-registers automatically when the agent reconnects.
Yes — the usage dashboard shows every request with the node that served it, token counts, latency, and your credit balance.
There is no formal SLA yet; the service is provided as-is per the Terms of Service. The agent reconnects automatically with exponential backoff if its connection drops.
Your node stops receiving requests until it reconnects — the agent retries automatically. If you run several nodes in a group, requests to the group endpoint are served by the nodes still online.
Two things: Ollama and the mealissa-llm-agent (about 440 KB), on Windows 10/11 (x64) or Ubuntu 20.04+/Debian/RHEL (x64). The agent installs a compatible Java runtime automatically if one isn't present. You'll also need a free license key. Full walkthrough: install guide.
If you can run Ollama, you can run Mealissa. Install Ollama, pull a model, download the agent, add your license key, run it — a few minutes total, with no networking configuration.
Ollama. Any model in the Ollama library works — Llama, Qwen, Gemma, Mistral, Phi, DeepSeek, and more — up to your plan's parameter size.
Yes. Mealissa implements the Anthropic Messages API — point the Claude SDK at your Mealissa endpoint and your local models answer in the format your code expects. See Compatible SDKs.
The Anthropic Claude SDK and the OpenAI SDK both work unchanged, in any language; plain HTTPS calls (for example with curl) work too.
Yes. Alongside the Anthropic Messages API, every node and group endpoint also implements the OpenAI Chat Completions API (/v1/chat/completions), including streaming with stream=true.
Any tool that lets you set a custom Anthropic- or OpenAI-compatible base URL can point at your Mealissa endpoint. There is no officially certified tool list yet.
Yes — request streaming through the OpenAI Chat Completions format with stream=true and tokens arrive as server-sent events. The Anthropic Messages endpoint currently returns complete responses.
Not yet. The API currently handles text conversations; tool calling is under consideration.
Yes — that's the core of the product. Your node gets one stable, globally reachable endpoint, encrypted in transit.
Yes. Any caller with a valid license key can send requests to your endpoint, and every request is authenticated individually.
Each running agent registers one model per node today. To serve several models, run additional nodes — they can share one group endpoint.
Yes, up to your plan's node count. Nodes are organized into groups, and requests to a group endpoint are distributed automatically across the group's online nodes.
The free Pico plan covers models up to 4b parameters, 1 node, 300 messages per month, and community support. There's also a 7-day free trial for models up to 14b — no credit card required. See all licensing plans.
Yes. Paid plans are priced by the maximum model parameter size your deployment runs; for larger or custom deployments, contact us.
Check the install guide or get in touch — we're happy to help.