Make scattered context searchable.
Vision OCR extracts text from screenshots. Local embeddings and SQLite organize the knowledge index so answers can point back to their sources.
Files + screenshots → local index
01 / Local intelligenceDesigned & built by Himanshu Jha
OmniDev
A working context. A native home.
A native macOS agent that searches files and screenshots with citations, edits code, and runs tests through scoped tools and approval prompts.
See the build01 / 03 · Playable recording
What to look for
From a request to inspectable tool activity.
Watch the native Mac app keep the conversation, working context and agent activity together. The recording shows the actual application; model waiting time is accelerated.
Native app recording · no narration · waiting time acceleratedOpen recording01 / The problemWhy this
Why this
needed building.
A developer’s useful context lives across files, screenshots and tools. Repeatedly pasting it into a chat loses the connection to the source—and a generated answer is still a long way from a safe edit.
I built a SwiftUI workspace around a local knowledge engine. The same context can ground a conversation, support a coding task, or be offered to another agent through MCP.
02 / Architecture
Follow the system.
The engineering behind the screens, in three stages.
Keep the source beside the answer.
The FastAPI engine serves source-grounded retrieval to the native app and exposes local capabilities through MCP. Ollama mode supports on-device inference.
Query → retrieval → cited response
Make a change with permission.
Scoped tools handle files and tests. Approval prompts and file snapshots make the boundary between a suggestion and an action visible.
Scoped tools → approval → restorable edits
03 / DecisionsThe choices
The choices
that matter.
One engine, several clients
The Mac interface and external MCP clients share the local engine instead of maintaining separate copies of the developer’s context.
Local is an explicit mode
Ollama supports local inference. Optional cloud providers have a different data path; the product does not describe every mode as offline.
Keep actions inspectable
Tool activity, approvals and restorable file snapshots matter as much as the generated answer.
04 / Evidence
Inspect it for yourself.
Real native-app recordings; local inference in Ollama mode.
Scope & limits
Explore the repository The recordings show the native app, with model waiting time accelerated. The separate web Agent Lab uses an in-memory fixture; it does not access a visitor’s filesystem. Planned integrations are labeled on the product page.
A useful agent needs a working context and a visible boundary around what it can change.
Next build / Agent-native interfacesGraft