Project management for Claude Code
Vibecode fast,
without losing track of the project.
Ovrsee keeps track of what you build with Claude Code: the plans you approved, the commits that carry them out, the tickets left to do, the site architecture and how it changes. Enough to know where you stand when you code at the agent's pace, and what to do next.
ovrsee
last scan · Aug 11 2026 · 0ca659f
Search, jump to…
⌘K
Views7
{{ v.label }}
{{ v.key }}
{{ v.count }}
Reorder, hide…
S
samuel
ovrsee
/
{{ viewTitle }}
{{ viewMeta }}
Reveal in Finder
Copy path
⌘⇧C
Compare two dates
Crawl
By tickets
By commits
Edit columns
Kanban
Epics
read from Graphify — 2026-08-14
Check live schema
All
Production
Without a reason
Terminal · claude
minimised — the session is still running
{{ statusRight }}
|
⌘K
interactive demo, rendered from the app code: click a view in the left column
How it works
Three commands, then the project keeps itself up to date.
01
Install once per project
$ pnpm ovrsee:install ./mon-projet
Sets up the Claude Code hooks and the post-commit hook. Nothing else to configure.
02
Map the application
$ pnpm ovrsee:crawl ./mon-projet
Playwright walks your routes and photographs every screen, dated and tied to the commit.
03
Read
$ pnpm electron
The app with its built-in terminal, or pnpm dev in a browser, ovrsee:brief as text.
From then on it is automatic: every plan approved in Claude Code is captured in ovrsee/plans/, and the post-commit hook attaches each commit to the active plan.
Seven views
Know where you stand without opening a single file.
Overview
Where the project stands, in one screen: open plans, tickets to do, repo health, branches, deployments.
Browser
Your app, its DevTools and an element picker inside the window. The selector pastes straight into Claude.
Product
The graph of crawled pages, screenshots included. Compare two dates: what changed shows.
History
What moved, week after week: every commit under the plan that explains it, activity as a histogram.
Board
The backlog as a kanban: epics and progress, WIP limits, one markdown file per ticket that Claude can write.
Data & Stack
Tables, columns and who uses them; dependencies with the reason written as a comment WHY: above the import.
With Claude Code
Claude stays in the loop, you keep control.
01
Built-in terminal
A real pty: login shell, claude running inside it, several sessions, dock at the bottom, on the side or full screen. The project commands (crawl, graph, Obsidian export) are one click away.
App-only, and over IPC: exposing it on a local socket would open it to every process on the same account.
02
Two skills included
ovrsee learns to read ovrsee/ : plans, pages, scans, screenshots and their reading traps. ovrsee-tickets learns to write board tickets, format and gestures included.
Installable from the setup screen, or with --skills.
03
MCP server
stdio, JSON-RPC 2.0, same interface as /api/* : Claude reads all of ovrsee/, writes only the tickets and board.json, runs no code.
Only projects in the registry are readable: a path that is not there is refused, even if it exists on disk.
The Data tab reads the code graph produced by Graphify : local AST parsing, every edge labelled. Ovrsee recomputes nothing: it detects the tool and shows its command.
The project, as files
Your backlog lives in your repository.
All the project tracking fits in one folder of your repository: it follows branches, reads back in any editor, exports to an Obsidian vault with its wikilinks. No proprietary format, no tool to sync.
Tracking follows branches: a plan travels with the branch that carries it
Every screenshot is dated and tied to the commit that produced it
Obsidian export with Dataview-queryable frontmatter
No code execution on the MCP side, no open socket
The code is on GitHub.
Electron + React + TypeScript, four production dependencies, no test framework. macOS arm64 and Windows x64 builds published on every tag.
unsigned binaries — macOS and Windows warn on first launch