An AI review inbox
for real work

Agents write code. You review, accept, or reject — with full context.

owlcoda is a handoff-first work harness. It manages the loop between AI execution and human decisions — across sessions, models, and people.

Get Started → View Source
owlcoda desktop app showing a dispatched task with context packet, agent assignment, and execution status
Early Public Release owlcoda v0.1 — Ready with Limitations
11 Supported 6 Partial 7 Not Yet
View all capabilities →

How it works

1

Describe the task

Define the goal and scope. owlcoda creates a structured task with context, accepted specs, and dispatch instructions.

2

Agent works in isolation

The assigned agent (Copilot, Claude, Codex) executes in an isolated git worktree. Your main branch stays clean.

3

Review the return

The agent returns a structured report: summary, files changed, key decisions, remaining risks, and a self-check.

4

Accept, reject, or continue

You make the final call. Accept to merge. Reject with reason. Requeue for another pass. Dispatch to a different agent.

What you see when work comes back

A three-column review surface. Tasks awaiting review on the left. Structured return document in the center. Files changed, self-check, and diff on the right.

owlcoda review surface showing returned tasks, structured return document with files changed and executor self-check

Structured returns, not chat logs

Every completed task produces a structured return document. Not a conversation transcript. Not a raw diff. A document designed for human review and decision-making.

Return Document impl-front-door-mvp · returned 
## Summary

Implemented front-door MVP with HTTP debug
adapter and Feishu adapter skeleton.
All required endpoints are in place.

## Files changed

- runtime/front-door/main.py
- runtime/front-door/adapters/http.py
- runtime/front-door/clients/session_store.py
- runtime/front-door/core/router.py
- runtime/front-door/Dockerfile

## Key decisions

- HTTP adapter has no auth (debug surface only)
- ExecutorProviderError on 4xx; JSON 502

## Executor self-check

- [x] All Python files parse cleanly
- [x] docker compose config validates
- [x] /healthz returns 200
- [x] POST /api/chat routes to executor
- [ ] End-to-end test (requires running stack)
Context Packet dispatched to copilot 
---
id: impl-front-door-mvp
type: context
task_type: coding
goal: "Implement minimal front-door service"
status: dispatched
---

## Accepted Specs

- spec-front-door-extract
- spec-session-store-public
- spec-entry-adapter

## Scope

In scope: main.py, adapters/http.py,
  clients/session_store.py, core/router.py

Out of scope: browser/search, rich media,
  session schema changes

Where owlcoda fits

owlcoda is not another AI coding IDE. It sits after the code is written — managing the review, handoff, and decision loop.

Cursor / Windsurf

AI-assisted coding IDEs. You write code with AI inline.

owlcoda doesn’t replace them. It takes over after they produce code — managing review, acceptance, and state continuity across sessions.

Copilot Workspace

A single-task AI environment for planning and implementing changes.

owlcoda manages multi-task, multi-agent, multi-round review lifecycles. Not one task at a time — a continuous work loop with handoff.

Devin

Autonomous AI software engineer. Strong on execution, less on human oversight.

owlcoda assumes you keep human decision authority. Accept, reject, requeue, and dispatch are first-class actions — not afterthoughts.

Local-first by design

Your code, your models, your machine. No cloud dependency. No data leaves unless you choose.

Local inference

Run models locally or use your own API keys. No cloud middleman between you and the model.

Git worktree isolation

Every agent task runs in an isolated git worktree. Your main branch stays untouched until you explicitly accept.

File-based state

All task state lives in .owlcoda/ — plain files you can inspect, version, and back up. No opaque database.

Inspectable protocol

Context packets, return documents, and dispatch records are structured text. Read them in any editor. Audit every decision.

Try it now

owlcoda is source-first. Clone the repo and start exploring.

git clone https://github.com/yeemio/owlcoda.git
cd owlcoda
Read the guide

Understand the concepts, entry paths, and how the harness works.

 See the walkthrough

Step-by-step paths through real review and execution workflows.

 Check capabilities

What works today, what’s partial, what’s coming next.

Good fit right now: Developers already using AI coding tools who want structured review and handoff for AI-generated work.

Not yet ready for: Non-technical users, production team deployments, or workflows requiring polished installers.

Open source, clear boundaries

Released under Apache-2.0. The code is yours to inspect, use, modify, and redistribute. The owlcoda name and brand assets remain project trademarks.