5
agents free per installation, no account
Why it exists
A working agent is more than a model called in a loop. It has to remember, survive a restart, call your tools and handle their failures, not lose work on a dropped connection, not run away in a loop, stay within budget and not leak provider keys. Teams usually build this subsystem themselves for months. AGENTBOX ships it as an image.
20
agents free when the project is identified
1
Docker Compose stack with an HTTP API
| What teams usually write themselves | What AGENTBOX does |
|---|---|
| The reasoning and tool-calling loop | One agent engine, the same in every DODVIR product |
| Memory between messages and across restarts | Stored in PostgreSQL; the agent is restored from the next event |
| Calling your tools | MCP: connect your server and its tools appear on the agents you grant them to |
| Deferred tasks and schedules | A scheduler the agent drives itself |
| Cron jobs and webhook handlers next to the agents | Automations: your script on a schedule, interval or webhook, without a model |
| Storing provider keys and secrets | Encrypted with the installation key, never returned outward |
| Limits, budget, loop protection | Cap on live agents, hourly and monthly spend cap per agent, step limit per turn |
| Recovery after a drop | Idempotent events, a read cursor, an SSE or WebSocket stream |
What an agent gets
Durable agents
A role, tool rights, memory and a lifecycle. An agent hibernates when its work is done and wakes up on the next event with its memory.
Your tools over MCP
Secrets for your API are stored encrypted and substituted at call time; the model never sees them.
Conversation and voice
Text events with idempotent delivery, a live SSE or WebSocket stream, voice calls and realtime speech-to-speech calls. Audio goes through your installation, never from the browser to the provider.
A desk per agent
Its own container with a persistent /work folder, a terminal and file tools. Your client folder is mounted read-only.
Knowledge, files, documents
Documents found by meaning, a library of skills granted through the role, file storage, PDFs read as text, images seen and drawn.
Your models
Each profile — reason, fast, embed, vision, image and others — is a direct connection to any OpenAI-compatible provider. You choose and pay for the model.
What it is not
- Not a library. It is a separate process with its own API, like PostgreSQL; you get no source code.
- Not a chat UI and not a permission system. One installation token controls it; users and access control stay in your product.
- Not model hosting. AGENTBOX reaches the model with your key.
Requirements
- Linux x86-64 with Docker and Docker Compose v2, and curl. There is no Windows or macOS build.
- A reasoning model with a context window of 250,000 tokens or more. Below that, the profile is rejected with a named reason instead of being silently truncated.
- Beta until 1.0: the API may change, and every change is in the changelog.
What leaves your server
- Requests to the model provider you configured — directly from your installation.
- A license check: the installation's identity and that it is active. Not your data and not conversations.
- Calls to your MCP servers, and whatever an agent reaches from its desk.
Questions
What is AGENTBOX?
A self-hosted runtime for AI agents by DODVIR. It runs as one Docker Compose stack with an HTTP API on your Linux server. Your backend creates agents, sends them events and reads their replies; AGENTBOX handles memory, tools, schedules, budgets, voice calls and recovery.
Is AGENTBOX free?
Yes, up to 5 agents per installation, without registration. An installation whose project can be identified — a real organization or product and the developer's email — gets 20 agents free. More agents need a paid license from the DODVIR account. Limits are set by the license server and may change; the current value is in GET /api/v1/status. You pay for your own models.
Which models does it work with?
Any OpenAI-compatible provider: each profile is a direct connection with its own address, key and model. The reasoning model needs a context window of at least 250,000 tokens.
How do I install it?
Run curl -fsSL https://dodvir.com/agentbox/install.sh | sh -s -- --dir ./agentbox on Linux with Docker and Compose v2. The installer gets a free license, writes .env with a fresh API token and starts the stack. Updating is one line in .env and docker compose up -d.
Can my coding agent set it up?
Yes. https://dodvir.com/agentbox/llms.txt is written for coding agents: install, update, identify the project, connect a model, create an agent, connect MCP tools.
Does it have a user interface or user accounts?
No. AGENTBOX is controlled by one installation token from your backend. The human-facing part and access control belong to your product.
Also by DODVIR
dodvir-1C: a decision system on top of 1C for mid-sized manufacturers in Russia and the CIS. Russian only.