Navire — Native Agent Vault, Interoperable Runtime Environment
Your agent is working from a copy of your work.
Give it the real thing.
Navire keeps your notes, boards, assets and canvases as plain files on your disk, then opens that folder to your agent over MCP and a CLI. It reads and writes the same vault you do — and everything it changes waits for your approval first.
WINDOWS X64 · MACOS ARM64 + X64 · EMAIL ONLY, NOTHING ELSE
An agent's work usually ends up as a wall of terminal output you scroll once and lose. Here it lands as things you can look at: a knowledge page, a card that moved, a proposal waiting on the board. The same vault, read two ways — a shell for the agent, an interface for you.
Your work does not live in a chat window.
Neither should the agents that do it.
The whole workspace runs from a shell
Boards, notes, checklists, knowledge pages, guides, plugin documents — every one of them has a command. It is the same surface the app uses, so an agent and a human drive the identical vault.
Batch dozens of operations in one round trip, or file the whole set as a reviewable changeset instead of executing it.
Your files on your filesystem. Not ours.
Knowledge pages are markdown. Boards, checklists and guides are readable JSON. Assets keep the names you dropped them under. Open the vault in your editor, version it with git, inspect it with ls.
SQLite is in there, but only as a catalogue and a search index outside the vault. Delete both and the app rebuilds them by walking the folder. Previous versions and deleted files sit beside the vault, so nothing goes away quietly.
Agents with names, roles and a chain of command
You author roles — a name, a remit, a tool list, routing rules — and Navire syncs them into your repository as files your agent CLI reads natively. Twenty ship in the default company, grouped into departments with their own boards and budgets.
You brief the chief. It reads the situation, proposes at most three moves, then routes each one by what the work is about — never by the role's name.
Build the part we did not build
The plugin API is the same one the app is built on, so a plugin is not a widget in a sandbox — it gets pages, documents, commands and an agent surface of its own. The SDK is a public package with docs covering the manifest, the ctx API and the permission model.
Scope it to one vault or the whole app, list your permissions, name the features you use. Whatever you ask for is what the install prompt shows the user, and nothing outside that list resolves at runtime.
From the catalogue, the installer checks the archive's sha256 and confirms the manifest id and version match what was claimed before anything touches disk. Sideloading a folder or zip prompts for the same permissions.
The SDK was written to be read by an agent as well as a person. Brief a plugin engineer role and it can scaffold, build and install one without leaving the shell.
Where everything sits
Navire makes no model calls. Your agent runs in the tool you already use, with its own credentials.
A desktop app on your machine, not a tab.
Named roles with remits, tools and routing.
Markdown and JSON in a folder you own.
Works with the editor, the git repo and the agent CLI you already use.
An MCP server and a CLI over the whole vault.
Extended by plugins, including ones agents write.
Questions
In an ordinary folder you chose. Knowledge pages are markdown, boards and checklists are readable JSON, assets keep their own names. You can open the same vault in your editor, in git or in a terminal without Navire running.
Yes. A filesystem watcher on each vault root picks up outside changes and the views re-list themselves. One caveat worth knowing: a document currently open in the app can lose an external edit, because autosave writes the app's copy back. Edit files that are not open and you are safe.
None of its own. Navire has no model integration, no keys and no inference calls. It exposes your vault over an MCP server and a CLI, and your agent runs in the tool you already use with its own credentials. Instruction files install into Claude Code and OpenCode today.
Nothing, through Navire. Whatever your agent CLI sends is between you and that vendor. Full-text search runs on SQLite FTS5 locally and semantic search on a dependency-free TF-IDF index, also locally, so nothing is uploaded to be indexed or embedded.
Two flags, both off by default, behind a first-run choice with equal-weight buttons. There is no endpoint compiled into the build, so even opting in sends nothing unless one is configured. If it ever does, the payload is a random install id, a session id, the app version and event names — never file contents, never paths.
It can run headless from your own scheduler — Navire has none of its own. What it does have is the review gate: agent edits to notes arrive as a proposal, the default trust level applies nothing automatically, applying is atomic, and every applied set writes an undo file beside it.
It works alone and it works shared. A collaboration server backs optional server vaults with accounts, invites and live change broadcast, and the document types built on the collaboration engine edit together in real time. Most of the app is built around one person's vault, so start there and add the server when you need it.
Nothing, and no account. It is free and unmetered, with no licence check or entitlement gate anywhere in the app. You point it at a folder and it indexes what is already there.
Windows x64 as a per-user NSIS installer, and macOS on both Apple Silicon and Intel as a dmg or zip. Auto-updates run on three channels: stable, beta and alpha. There is no Linux build.
Yes, and against the same API the app itself is built on. The SDK is a public versioned package installable by URL, with docs covering the manifest, the ctx API and the permission model. Declare what you need, install from the catalogue with the checksum verified, or sideload a folder or zip while you work on it.
It goes into a database we run, and it gets one email — the day you can download the app. No newsletter, no drip sequence, nobody else gets the list. We keep the address, whatever campaign tag brought you here, the country your request came from and your browser's user-agent string; your IP address is hashed for rate limiting and never stored in the clear. Ask and it is deleted.
Point it at a folder and see
No account, no key, nothing to configure before it opens. Pick a folder and it indexes what is already there. Leave an address and we will tell you the day you can.
WINDOWS AND MACOS · ONE EMAIL WHEN IT OPENS · NO TRACKING