Skip to content
MIRA docs

What is MIRA?

MIRA (Multi-tasking Intelligent Responsive Assistant) is a self-hosted personal AI agent. You run it on your own machine; it talks to whichever LLM provider you configure — Anthropic, OpenAI, OpenRouter, DeepSeek, Gemini, or a local / OpenAI-compatible server. Your conversations, memory, and data stay on your hardware.

If you’ve used ChatGPT or Claude, MIRA will feel familiar to chat with — but it does several things a chat app doesn’t.

How MIRA is different from a chat app

Section titled “How MIRA is different from a chat app”

It reaches you on real channels. MIRA isn’t only a web tab. It can message you on Signal and Telegram, send and receive email, post to Discord, Matrix, Slack, or WhatsApp, and push browser/phone notifications. It can also start conversations — a proactive check-in, a morning briefing — not just answer when asked.

It remembers. MIRA keeps a growing memory of facts about you and a personal wiki of longer notes, and uses both to personalise replies across sessions and channels. Tell it something once and it carries forward.

It acts, through tools. Beyond chat, MIRA runs tools: web search and fetch, sandboxed code execution, your calendar, PDF extraction, scheduled automations, and — through MCP (Model Context Protocol) — any external tool server you connect, from browser automation to GitHub to your own scripts.

It speaks. Built-in text-to-speech and speech-to-text mean replies and check-ins can arrive as voice notes, and you can talk back.

It serves more than one person. A single MIRA instance can host several people, each with their own account, memory, and settings, plus an admin who manages the server.

The shape of the system

Section titled “The shape of the system”

MIRA runs as a single gateway process that hosts the agent, the web/HTTP API, the channel pollers, and the schedulers. There’s no database server or message broker to stand up — it’s one binary plus a data directory.

┌──────────────────────────────────────────────────────────────┐
│  Channels:  Web · Telegram · Signal · Email · Discord ·       │
│             Matrix · WhatsApp · Slack · Push · (your own)     │
└───────────────────────────┬──────────────────────────────────┘
                            ▼
┌──────────────────────────────────────────────────────────────┐
│  Gateway (one process)                                        │
│    Agent core ── multi-LLM router · Memory + Wiki ·           │
│                  Tools + MCP host · Automations · Auth         │
└───────────────────────────┬──────────────────────────────────┘
                            ▼
         Data directory (~/.mira/data): your DBs, wiki,
         artifacts, and voice models — all on your disk

Settings come in two kinds:

  • Operator / global settings live in mira_config.json — providers, channels, security, voice, and so on. These are admin-managed and apply server-wide.
  • Per-user settings belong to each person — their voice preferences, companion configuration, channel accounts, connected MCP servers, and profile.

Your data lives in a data directory (~/.mira/data by default): the auth/user database, the memory database, the wiki, the companion database, artifacts, and downloaded voice models. You can put it anywhere — a backed-up volume or an external disk — by choosing it during mira setup, setting data_dir in the config, or passing --data-dir. Backing up that one directory backs up your whole MIRA.

Asking MIRA about itself

Section titled “Asking MIRA about itself”

MIRA ships with its own documentation built into the binary. You can ask it things like “what can you do?”, “how do I enable companion check-ins?”, or “what does the agent.max_tool_rounds setting do?” and it will answer from that built-in reference. For the current value of a setting — and to change one — MIRA respects your access level: you can see and change your own settings, operator/global settings are admin-only, and secrets are never shown.

Next steps

Section titled “Next steps”