Sessions stay siloed
Each supported coding agent keeps its own session and transcript, so no single record shows the work in flight.
Oko core for macOS
Understand what your AI-native team is working on, and why
Oko synchronizes, analyses, and interprets transcript-backed coding-agent work. It is one Rust program, oko, with its read-only MCP server and one resident service, on macOS 14 or newer.
- macOS 14 or newer
- One program: oko
- Loopback goal inference
- Autonomy off by default
Your agent work is scattered across tools
Coding-agent output lives in vendor sessions, local processes, transcripts, Git repositories, worktrees, goals, and routing state. Reconstructing what people and agents are doing, and why, takes manual archaeology.
Processes and repositories drift apart
Local agent processes, Git repositories, and worktrees each hold part of the state that explains a change.
Goals and routing lack context
Goal and routing state sits apart from the transcript record, so the reason behind the work is hard to recover.
How Oko turns scattered sessions into one searchable record
Oko core owns discovery, masked transcript import, normalization, and a disposable search projection. The interfaces stay small and inspectable.
Discover local agents
Oko owns discovery of supported local coding-agent processes running on the Mac.
Import with masking
Oko imports masked, provider-neutral Transcript Lake exports.
Normalize in oko
oko normalizes transcript, session, process, Git, synchronization, search, and routing logic.
Project for search
The catalogue in the shared Oko database powers search, and oko mcp serve reads it without writing.
Reason over loopback
The qualified goal-lifecycle model is served on loopback only for the command that classifies and exits with it, so transcript text stays on the Mac during inference.
Every interface has a documented boundary
The product contract states what each interface owns and what it refuses to do. Oko core and Oko Desktop stay visibly distinct.
| Interface | What it owns | Boundary |
|---|---|---|
| oko | Catalogue, goals, telemetry, calendar, Slack, sync, routing, diagnostics | One Rust program |
| oko serve | The one resident process; owns the PTY broker | Cadenced work runs as Stado routines |
| oko mcp serve | Read-only catalogue and team-view server | Never indexes or writes |
| OkoKit | Decode-only facade for Oko Desktop | Swift package; decides nothing |
| Goal lifecycle | Qualified model served on loopback for the classifying command only | Transcript text stays on the Mac |
Questions before you build on the core
Does Oko core ship the native desktop app?
No. The native macOS application bundle, SwiftUI interface, lifecycle, and Sparkle updates belong to Oko Desktop, a related product.
Where does goal inference run?
On the Mac: the command that classifies serves the qualified goal-lifecycle model on loopback for its own run and stops it afterwards, so transcript text stays on the Mac during inference.
Can agents act autonomously by default?
No. Autonomous execution is disabled until an explicit persisted policy enables it, and accepted goal completion is receipt-gated.
Which platforms and toolchains are supported?
Oko core currently supports macOS 14 or newer. Source builds require a Rust toolchain.
Read the boundary, then build on the core
Oko helps teams understand interactions with AI by synchronizing, analysing, and interpreting transcript-backed work. Start with the product contract, then inspect oko, oko serve, and oko mcp serve. Requires macOS 14 or newer.