miton

In development. The beta will be a downloadable desktop app with automatic updates. Access is not open yet.Request an invitation.

Reference · 3 min read

Architecture concepts

How Miton works in plain English — one project, specialised workspaces, shared goals, memory and permissions — before the engineering deep dives.

Product stage
Pre-release development
Last reviewed
2026-07-23
Successful result
The change is visible in the active project.

Who this is for: curious users who want the product’s big picture before choosing a task guide.

How Miton works (60-second version)

Miton is a chat-first AI workspace on your machine. You describe what you want; Miton plans, writes, runs tools, and shows results in the conversation.

  • One chat carries context across everything you do in a project.
  • Four specialised workspaces — Writer, Canvas, Code, and Automation — organise the same project for different kinds of work. Chat remains available throughout.
  • Your keys, your data — API keys live in the OS keychain; projects live in folders you control.
  • You choose the guardrails — workspace trust plus a per-session steering mode (Confirm is the safe default) decide what runs without asking.

The work changes shape. The project does not.

The sections below explain the complete public model without requiring source or contributor documentation.

The specialised workspaces

Surface Best for
Writer Long-form text, citations, tracked changes, export
Canvas No-code apps — pages, components, publish
Code Software — editor, terminal, git, search
Automation Recurring workflows — schedules, webhooks, visual graphs

All four share one agent engine, one memory store, one model router, one permission layer, and one undo stack. Switching surface changes defaults and tools, not the project underneath.

Surface how-tos: handbook index.

The agent engine

The engine is a loop: read your prompt → call the model → parse the response → run tools → render results in chat. It is shared across surfaces and extended with skills, sub-agents, and custom commands.

Memory

Memory stacks in layers that compose into the agent’s context each turn:

Layer What it holds
Session log Messages in the active chat
Session archive Past sessions for this project
Project memory MITON.md, .miton/rules/, plan, goals
Cross-session memory Preferences and corrections you have taught Miton

Smart routing

Each turn, Miton picks a model using your routing profile (free through premium), task complexity, budget state, and quality preference. Change profile per session in the model chip or with /profile.

Permissions and safety

Trust (per folder) and steering mode (per session) compose to decide what the agent may read, write, and run. A safety layer scans prompts and tool output, redacts sensitive egress, and enforces gates before actions execute.

Reference: permissions taxonomy.

Cross-surface flows

Evidence that the surfaces share one engine:

Flow What it does
Promote to project Turn an Canvas build into a Code project
Visualise in canvas Open a Code component in the designer canvas
Convert to workflow Turn a one-off action into an Automation workflow
Generate docs Produce Writer documents from Code or designer work
Workflow from prompt Sketch a workflow in Automation from natural language

These are visible buttons in the UI — not hidden power features.

What this page does not cover

  • Step-by-step usage — use the task guides linked from Docs.
  • Internal package names and contributor implementation records — these are not part of the public product vocabulary.

See also

Something failed? Keep the exact error, confirm the active project and model connection, then use Troubleshooting.