Product

Agents that work for you, under rules you set.

People ask an assistant for work in plain words. The assistant does it through agents your organization installed, and the platform checks every step.

From a request to the work

The model proposes; the platform decides.

  1. 1

    You ask

    In a chat, typed or spoken. A schedule or another system, through a webhook, can start work too.

  2. 2

    The assistant chooses

    The model picks which agent to call and with what. It authorizes nothing.

  3. 3

    The platform checks

    The call must be one the agent declared, allowed by your organization's rules, and within the chat's trust. Above that trust, the person approves.

  4. 4

    The agent works

    Its program runs in a sandbox, reads and keeps only what it declared, and the call goes on the audit trail.

What you control

Every control is set before an agent runs.

Enforced by the platform on every call. Nothing depends on the agent behaving well.

Which agents exist

Administrators add the catalogs they trust and install one exact version after reading its declaration. Its files are checked against their digests, and it is started once with no network to prove it is what it says, before anyone uses it.

Who may use what

People belong to groups, and groups to policies. A rule allows, denies or asks, for people and for agents' actions alike. A denial always wins, and nothing is allowed by default.

How far each action reaches

Every action declares a level: reading, changing, wider changes, or acting outside the platform. One above the chat's trust waits for a person's approval, which can be remembered for that agent.

Credentials and secrets

Stored encrypted, never shown to the model. A secret reaches only an agent its record lists, during a call. For a connected account the agent gets a fresh access token, never the password.

Where agents connect

Only to the hosts an agent declared, or ones a person opened for a single call. Public addresses only, and never a site your organization blocked.

The audit trail

Every call, approval and refusal, by person and by time. Written by the platform, not the agent, and never edited.

Every day

What people and administrators use it for.

Chats

An assistant in each chat, with its own agents, model and trust. Cards ask for an approval, a choice or a sign-in, right where the work is.

Records you own

Data, files and secrets agents keep for you, each with an owner and a list of the agents that may reach it. Shared or handed over by hand.

Work that repeats

Schedules run a chat's work on a cadence and report when there is something to say. Webhooks let another system start work, at a trust you cap.

Notifications

Web push, email, or addresses you choose, such as Slack or Teams. An organization can make them say only that something happened.

Speech

Dictate and listen, with speech models on your own machine or a model connection you choose. Off until you turn it on.

The assistant's own tools

Memories and skills the assistant keeps for a person, and the MCP servers they add, which an organization can turn off.

Several organizations

One installation can host several organizations, each kept apart, each able to hold its chats on runtimes of its own.

Scripts and systems

Everything the web client does is an API call. Scripts call it with an API key that expires.

Built to be checked

Small parts, each doing one thing.

Every call between them is HTTPS with mutual TLS, by certificates from your deployment's own roots.

API server

The platform's data and who may do what with it: people, agents, records, chats, settings and the trail. It runs no agent's code and calls no model.

Runtime

Holds the active chats: their agents' processes, the check on every call, the way out to the network, and the assistant that thinks in each. It keeps nothing.

Sandbox

Runs each agent's program in a jail of its own: its own files, its own network, its own memory limit. It knows programs, never chats.

Issuer and speech

The issuer certifies every other part. Speech turns voice to text and back with models on your machine.

To see each part and why it is what it is, ask us for a demo at ahmed@decentai.io.

Try it

Run DecentAI on your own computer.

Write to us for access or a demo. One command sets it up with Docker; add a model, add an agent catalog, and ask.