Product · AI coding agents you supervise
Agentic Taskboard
A project board where every task is worked on by a real Claude Code agent. The agent writes a plan, you approve it, it does the work in an isolated copy of your repository, the checks run, a second agent reviews the diff, and nothing reaches your code until you accept it.
- Built on
- Claude Code (headless sessions)
- Stack
- FastAPI · SQLite · vanilla JS · PWA
- Status
- Early access · used daily by its builder
- Made by
- Raed Ouiriemmi, Tunisia
The problem
Coding agents are good enough to do real work, but running several of them at once is hard to keep under control: they edit files you did not expect, they finish while you are away, and reviewing their output means digging through terminals and diffs. Agentic Taskboard puts every agent session on a board, with one rule: the human decides at every step.
How it works
- Plan. You write a task and pick an agent (frontend, backend, QA, reviewer…). The agent reads the code in read-only mode and writes a plan.
- Approve. You approve the plan, edit it, or reject it with feedback. The same session continues, so the agent keeps its context.
- Work in isolation. In safe mode each task gets its own git worktree and branch. Your project folder does not change while the agent works.
- Check and review. The project's checks (tests, build) run after every change; on failure the agent gets one automatic fix round. An optional reviewer agent reads the diff and lists bugs, risks and nits.
- Accept or discard. An evidence card shows the checks, the size of the change, merge conflicts and the review. Accept merges the branch, Discard throws it away.
What it does
- Overview: everything waiting for your decision, plans first, with one-click Approve or Accept.
- Timeline: agents × weeks, drag a card to reschedule, project health and what is late.
- Scheduling: tasks can start on a day and time, repeat daily or weekly, or wait for another task.
- Agent isolation: each run gets an allow-list of tools and deny rules for secrets (SSH keys, cloud credentials, the board's own data). Anything blocked is reported, with “allow once”.
- Line comments on the diff, sent back to the same agent session.
- Telegram alerts for plans to approve and finished work, with Approve / Accept / Discard buttons.
- Secure access: password and two-factor login; installable as an app on desktop and phone.
Why Claude
Every agent is a real Claude Code session started in headless mode in the project folder, resumed between the plan, the work and each round of feedback. The reviewer runs as a separate read-only session. Next on the roadmap is running agents directly on the Claude API, so a team can use the board without each member's own Claude subscription, and sharing one board between several people.
Early access and contact
Agentic Taskboard is in early access. To try it or talk about it: Raed Ouiriemmi · ouiriemmiraed@gmail.com