Grounded agents that don't just talk — they act.

Q. What is Radioso?

Radioso is a self-hosted context platform for grounded assistants. You upload your documents and Radioso parses, chunks, and embeds them into Postgres with pgvector1. From there, one backend powers a web chat, a REST API, a TypeScript SDK, a website embed, and an MCP server — so the same grounded answers reach your end users and your tools2. The data and the model keys stay in your stack3.

Sources1docs/retrieval.md· § pgvector2docs/architecture.md· § Surfaces3docs/deployment.md· § Self-hosting
Answers by Radioso

Self-hosted

Runs on your own infrastructure

Bring your own LLM

Your keys, no inference markup

Your data stays put

Everything in your Postgres

Open standards

MCP, REST, and SDK included

Inside Radioso.

Assistant

one turn loop · steer, act, follow through

Directives

steer
confirm before acting
stay concise
+ your rules

Standing condition → action rules. Shape behavior, no new capability.

Skills

act

retrieval.answer

ground answers in your docs

Your own

register a capability

retrieval draws on your docs → Postgres + pgvector

Routines

guide

Contact a human

collect → confirm → send

Your own

author a journey

Stateful multi-step journeys, resumed across turns.

LLM
Grounded answer + citations·async actions (outbox)
Web chatPublic chatWebsite embedREST APITypeScript SDKMCP server

Ask from any surface — the answer returns there.

Talks to your users and takes actions you need.

It does the work, not just the talking.

A bot that can only talk is a glorified search box. Radioso agents follow the rules you set, run the procedures you author, and take real action — then show you exactly why.

Directives

steer

Standing rules that shape how the agent behaves. "When the customer sounds anxious, slow down and confirm before doing anything." Write it once; the agent applies it whenever it fits, on every surface, in every language. A directive steers — it never acts on its own.

Routines

guide

Multi-step flows the agent runs across turns — a return request, an onboarding check, "collect these three things and file a ticket." Write the steps in plain language, drop in a chip to collect a value, call a tool, or branch. Radioso compiles it and the engine resumes it turn to turn. No canvas. No redeploy. Validate, publish, done.

Act & hand off

act

When the agent should stop talking and do something, it does: call a tool, fire a webhook when a routine finishes, or hand the conversation to a real person. Every step it takes lands in the same turn trace as the answers — so when it does the wrong thing, you can see why.

Working assistant in under five minutes.

The bootstrap script prompts for your AI provider credentials, generates secrets, and starts the full stack with Docker Compose. Register, upload a document, ask a question.

~/radioso
# requires Node 24 + Docker Desktop
./run-dev.sh

# then open
# http://localhost:3000   web app
# http://localhost:8080   API