UnreliableWitness/Shellie

One work item, one worktree, one set of agent terminals, and a bill you can send.
Shellie turns an Azure DevOps work item into an isolated workspace: its own git worktree and branch, Claude Code,
Copilot and shell terminals, and an exact count of the AI cost and time the task took.
Download for Windows ·
macOS ·
Features ·
Getting started ·
.shellie.yml ·
Development
## Why Shellie
Working with AI agents means running several tasks at once, each in its own checkout, each with an agent burning
tokens. Shellie gives every task its own room and keeps the books:
No more branch juggling. Every task is a git worktree on its own branch, set up and ready before you type.
Agents with context. Claude Code starts knowing the work item it is working on.
Cost you can invoice. Token usage is captured per task and priced at the official list prices in EUR, on
consumption-based licences, so it can go on the customer's bill.
Azure DevOps stays current. Work items move to In Progress, draft PRs are created and linked, and time and cost
can be written back.
## How it works
mermaid
flowchart LR
A["Work item in the sprint"] -->|Start| B["Worktree + branch setup commands"]
B --> C["Terminals Claude Code · Copilot · shell"]
C --> D["Review changes comment, send to agent"]
D -->|Finish| E["Draft PR cost + time booked"]
E --> F["Done reopen any time"]
1. Pick a work item on Home or the Sprint page and press Start.
2. Shellie creates the worktree and branch, copies what the repo needs, moves the work item to In Progress and assigns
it to you.
3. Work in the task's terminals. Setup commands such as npm ci run in their own tab, so you can start at once.
4. Review the changes in a diff view, leave comments on lines and send them to an agent.
5. Finish: Shellie pushes, opens a draft pull request linked to the work item, books the time and cleans up the
worktree. Finished tasks keep their summary and can be reopened.
## A look around
Home : your tasks, pull requests and sprints across organizations, with the week's cost and time.
Sprint : the current sprint in backlog order, time spent against the estimate, and Start on every row.
Changes : everything the task changed, in the Monaco diff editor, with line comments for the agent.
Usage : cost per model, tokens, your time against the agents', and the last six hours.
Done : what a finished task took: commits, the pull request, the AI cost and the time booked. Reopen brings it back.
Screenshots show a demo project with demo usage data.
## Features
### Task workspaces
A git worktree and branch per task, named from the work item (feature/1234-fix-login)
Per-repo setup from .shellie.yml, with node_modules cloned instantly on macOS when the lockfile matches
Tasks that span several repos: detected from what the agent edits, with one PR per repo
Focus mode, a tab per terminal, and terminals that come back after a restart
### Agents in the terminal
Claude Code, GitHub Copilot CLI, OpenCode and your shell, side by side
Claude Code gets the work item as context and keeps its session across relaunches
Clickable path:line and links, Send to a terminal, and copy and paste like VS Code
Agent packages installed per harness when the repo has an apm.yml (only for the harnesses its targets: lists, when it has one)