Traction
Verified progress only. We publish real milestones — never fabricated metrics.
This page is built to be checked. Everything on it is either a dated release you can match against the changelog, or a capability you can confirm by installing Git Lora.
There are no user counts here yet, and that is a deliberate choice rather than a gap. The section at the bottom says exactly which numbers will appear, how they will be counted, and what has to be true before we publish them.
How the product got here
- 1.0May 2026
Foundation
Local-first workspace. A project is an ordinary folder; Markdown editor, file tree, and the first canonical memory index.
- 1.2May 2026
Graph and entities
Entity extraction from your own files, a relationship graph, and fact tiers — canon, working, sandbox — with contradictions surfaced rather than averaged.
- 1.5June 2026
Companion and MCP
A conversation grounded in canonical memory, private companion memory, and an MCP server so external clients can query a project without copying it.
- 2.0June 2026
Agent OS and the proposal rule
A persistent background runtime, and the rule that no AI edit is ever written straight to a file — every one becomes a reviewable proposal.
- 2.3July 2026
Autonomous Engine
Plan, execute, checkpoint, recover, QC, continue. Long work that survives a failure halfway through.
- 2.6July 2026
Finalizer and branches
End-to-end assembly of a long artifact under constraints set once, and alternative development branches inside a project.
- 2.8August 2026
Sync and languages
Optional sync to storage you control, three interface languages, and import of existing Markdown vaults.
- 3.0August 2026
Service layer
Installation registry, signed and verified updates, in-app support, and a control panel with separated staff roles — without giving up working offline.
What you can verify today
It runs without us
Disconnect from the network. The editor, memory, graph and agents keep working, including startup, which never waits on our servers.
Your files stay files
A project is a folder of Markdown and code. Nothing is converted into a private format, and Git works over it normally.
Nothing is overwritten silently
Every AI edit is stored as a proposal, shown as a diff and passed through a quality gate that can refuse it. This cannot be switched off.
Contradictions surface
When two canonical facts disagree, the conflict is shown rather than quietly resolved to an average.
No hardware fingerprinting
The installation identifier is a random value made on your machine. No MAC address, machine name or serial number is collected.
Updates are verified
Publisher, signature and SHA-256 are all checked before an installer runs. A build without a checksum is never offered.
Which numbers will appear, and how
Download counts and active-user figures are easy to print and impossible for a reader to audit. Here is what we intend to publish, with the definition attached — because a number without its definition is decoration.
- Registered installations
- Count of distinct installation identifiers that completed registration. Reinstalling produces a new identifier, so this over-counts people and we will say so next to the figure rather than quietly presenting it as users.
- Installations seen in the last 30 days
- Registrations whose last contact falls inside a rolling 30-day window. This is liveness, not engagement: it says the app checked for an update, not that anyone wrote anything.
- Distribution by platform and channel
- Windows / macOS / Linux and stable / beta. Coarse by design — it exists to decide which installers to build.
- Support load and resolution time
- Tickets opened and median time to first reply, taken from the same records the support team works in.
- What we will not publish
- Anything derived from project content, and any figure we cannot define in one sentence. We do not have project content, so metrics about what people write are not available to us at any price.