Local-first AI coding harness
AI coding agent control room.
Run Codex, Claude Code, OpenCode, local repository context, swarm orchestration, terminal output, and diff review from one red desktop console.
agent.lobrecs.tech
$ lobrecs-agent run --repo ./product
context.scan: 2,184 files indexed
router.pick: codex high reasoning
swarm.dispatch: planner -> implementer -> reviewer
terminal.capture: tests, diffs, failures
quality.gate: patch applied, review ready
Control room
A desktop IDE for agent work that has to be inspectable.
Lobrecs Agent keeps the prompt, local command stream, repository context, code changes, and review state in one place. The app is a harness around the tools developers already use instead of a black box that hides what changed.
Session timeline
Turn-based messages, commands, approvals, edited files, and completion evidence stay together.
Repository context
Project memory and semantic code context enrich prompts without turning private work into cloud state.
Swarm orchestration
Planner, implementer, reviewer, and debugger flows can run against one repository without losing thread state.
Review-ready diffs
Completed patches are applied by the main process and surfaced in the renderer as evidence for review.
Features
Everything is organized around visible agent work.
Each capability follows the same pattern: gather local context, choose the right agent, stream the evidence, apply completed diffs, and leave the user with reviewable output.
Model routing
Small fixes, medium refactors, and high-reasoning work can be routed to different CLIs and models based on task complexity.
Local context scan
The app indexes repository files and project memory so prompts include the codebase facts that matter for the current task.
Terminal capture
Commands, failures, tests, and repair attempts are preserved as artifacts instead of disappearing into a background process.
Quality gate
Completed sessions can run focused verification and show whether the patch is ready, failed, or needs another pass.
Diff review
File edits are applied locally by the main process and surfaced in the UI with enough context to inspect before shipping.
PR handoff
Git-aware helpers prepare commits, branch state, and pull request context so finished agent work can move into normal review.
Workflow
From prompt to verified patch without hiding the machine.
The same run structure appears throughout the app: context first, agent selection second, observable execution third, review last.
-
01
Select the repo
Point Lobrecs Agent at a local checkout and keep execution on the machine.
-
02
Dispatch the right agent
Route easy, medium, and complex tasks to the model and CLI that fits the job.
-
03
Watch the evidence
Commands, terminal output, approvals, and file edits become visible artifacts.
-
04
Apply completed diffs
When the agent finishes, the main process applies the patch locally and keeps the review UI focused on inspection.
-
05
Review and ship
Use the diff, quality gates, and PR helpers to move from completed work to delivery.
Open repository
Built in public workflow style, ready for contributors.
Contributions are most useful when they preserve the local-first security model: no API keys in storage, narrow preload APIs, visible terminal evidence, and deterministic review states.
repo
oliveirabalsa/lobrecs-agent
Electron, React, TypeScript, SQLite, xterm.js, Monaco, and local agent CLIs working as one desktop control room.
Clone
git clone https://github.com/oliveirabalsa/lobrecs-agent.git
Install
npm install prepares Electron, native modules, and the renderer toolchain.
Run locally
npm run dev opens the desktop app with the local repository workflow enabled.
Verify
npm test runs the Vitest suite before a pull request or release candidate.
Desktop app
Built for people who want agent speed without losing local control.
Download packaged builds from GitHub releases or run the project from source when contributing.