Send less useless context
Tool output is bounded, repeated payloads are removed and older turns are condensed for the task at hand.
A workspace for the work after the prompt
Ship the whole project. Not just the prompt.
Models can write a useful answer. Miton turns that answer into a coherent project—with interfaces, documents, automations, code, goals, costs and the evidence needed to ship.
01 / Finish something real
The answer is the beginning.
A serious project does not end when a model stops generating. It needs decisions that survive the next conversation, a front end that matches the intent, a back end that works, documentation people can use and an artefact somebody can actually run.
Miton keeps Chat beside the work while specialised surfaces split out when they are useful and merge back into the same project. The development build already carries active project context and goals across the workspace. The complete delivery loop is the explicitly labelled v1 programme—not a present-tense promise smuggled into a hero.
02 / Built, not borrowed
Not a fork. Not a wrapper. Not somebody else’s interaction model with a new logo.
Miton is its own Tauri application, agent runtime and project system. That ground-up choice lets Chat stay constant while Canvas, Writer, Automation and Code organise themselves around the task. It also means model routing, cost control, memory and safety can share one project contract instead of being bolted onto an editor extension.
Read the architectural argument →One conversation and one plan follow Atlas Migration into every workspace. The selected worktree is part of the project context, not a separate Chat.
03 / Model intelligence
Use the model the work deserves.
Miton scores eligible models for each task while respecting provider availability, user routing posture and model capabilities.
A cheap local model can classify a simple request. A difficult architecture decision may deserve a stronger hosted model. Named groups can narrow the eligible pool to local, inexpensive or frontier models; Miton then scores the connected candidates, respects your provider boundaries and explains the cache, capability and cost signals behind the choice.
Named model groups narrow the eligible pool; Miton still scores the connected candidates and explains the cache, capability and cost signals behind the choice.
04 / Token economics
Frontier tokens are a resource, not a ritual.
Keep the warm prefix stable
Cache-friendly context stays in a predictable order so supported providers can reuse work instead of charging to read it again.
Route every sensible chance
Simple work does not automatically consume the most expensive model connected to the workspace.
Show the meter
Costs, model use and project totals stay visible, with caps available before surprise spend becomes a post-mortem.
05 / Local models, first-class
Local is a capability boundary, not a consolation prize.
A local model should get the same careful harness as a hosted one: capability checks, context shaped to its limits and tools it can actually use. It should not be flattered into doing intensive work it cannot do well when a better model is available and permitted.
Miton treats local and hosted models as eligible components of one system. You choose the boundaries; routing chooses within them.
06 / One project, every workspace
The work changes shape. The project does not.
Canvas, Writer, Automation and Code are projections of one project—not destinations for copied output. The conversation and Plan stay shared. The selected worktree decides what Canvas and Code show, and comparison with main remains explicit.
01
Chat
02
Canvas
03
Writer
04
Automation
05
Code
Canvas follows the selected worktree. Source remains canonical; App Kit alignment is offered rather than forced, and changes can be compared with main.
07 / Safety before agency
Power without an audit trail is just anxiety.
Miton checks untrusted instructions, isolates secrets, applies permission policy before tools run and records supported outbound activity. Safety is spread across the surfaces where decisions happen, then made inspectable in one place.
Inspect the trust model →A current deterministic capture of the production Safety and privacy controls. This is literal interface evidence, not a product render.
08 / Implemented for beta
What works in the development build.
Present tense means verified behaviour, reviewed on 23 July 2026. It does not mean a public download is available yet.
Beta release gate: The public beta will be delivered as a signed desktop download with updates handled inside Miton. It is not publicly available yet and will not use build-from-source as its release path.
- 01
Built from the ground up
Miton is its own Tauri and React application, not a fork of VS Code or a wrapper around another agent harness.
- 02
One project across the workspace
Chat, Canvas, Writer, Automation and Code operate on the active project and share project goals, knowledge and conversation context.
- 03
One conversation and revisioned project state
Workspace changes keep the active project conversation, plan and supported artefacts together; the selected worktree is canonical and conflicts are explicit rather than last-writer-wins.
- 04
Canvas follows the selected worktree
Canvas projects the front end from the selected worktree, keeps source canonical and marks constructs outside its supported visual representation as code-owned.
- 05
App Kit design floor and optional realignment
Miton-scaffolded front ends use the App Kit design floor; imported front ends receive a read-only alignment check and an explicit realign-or-keep choice.
- 06
Architecture Map v1
Canvas indexes a project into an honest front-end, back-end, data and infrastructure topology list, including a recovery path when a front end cannot be projected.
- 07
Manuscript routing and starter
Novel, book, chapter and manuscript requests route to Writer, whose eighteenth starter provides a book bible, chapter outline and draft chapter structure.
- 08
Task-aware model routing
Miton scores eligible models for each task while respecting provider availability, user routing posture and model capabilities.
- 09
Named model-routing groups
Users can create named allowlists such as local only or cheap coding and activate one globally before Miton scores eligible connected models.
- 10
Inspectable routing decisions
Routing choices expose the selected model and supporting cost, capability and policy signals instead of hiding the decision.
- 11
Token economics as a system concern
Miton condenses context, preserves cache-friendly prefixes, records spend and supports limits so frontier tokens are used deliberately.
- 12
Local models are first-class
Local providers participate in the same capability checks and routing controls as hosted providers without pretending every model suits every task.
- 13
Workspace licence separated from model spend
Miton connects to user-selected hosted and local providers rather than reselling a bundled token allowance.
- 14
Safety before agency
Miton combines permission policy, injection checks, secret isolation, audit events and recoverable project actions.
- 15
Inspectable network egress
The desktop records supported outbound activity in one ledger and verifies both WebView and Rust request surfaces.
09 / The v1 programme
The complete real-software loop.
These are intended v1 outcomes. They remain outside beta present tense until their claim gates are green.
Equivalent Windows, Linux and macOS outcomes
The v1 programme targets the same core features and project outcomes across Windows, Linux and macOS.
Documentation produced with the project
The v1 programme joins project knowledge, Writer and delivery workflows into a maintained documentation suite.
Front end and back end in one delivery loop
The v1 programme extends Canvas and Code from the shipped supported projection into complete framework write-through, verification and repository delivery.
A shippable project artifact
The v1 programme closes the loop from project intent to a verified, packaged artifact with an inspectable hand-off.
10 / How Miton is made
Built heavily with LLMs. Owned by people.
Miton does not hide its own use of models. LLMs contribute substantially to its code, tests, documentation and analysis. Foundry Seven sets the direction, decides what is accepted and remains accountable for what ships.
That leverage lets a small team build more. It does not replace verification, security response, maintenance or the need for somebody to keep showing up.
Read the candid licensing argument →11 / Pay for the workplace, not the model
Your models will change. The system that gets good work from them should compound.
Miton does not build its value around a token bundle. Bring hosted keys or local models. Miton tracks the changing model field, improves the harness and helps every cent do more useful work.