06:00:00 schedule Tidy README fired (0 6 * * 1-5) 06:00:04 workspace greeter cloned at 3039773 06:00:05 run kv58aa5b, agent mode 06:01:41 run 1 file changed, $0.15 06:01:41 gate waiting for approval
// claude on a schedule
Claude keeps the watch. You keep the helm.
qubichelm runs Claude against your projects and repositories on a schedule you set. It reads the backlog, does the work in an isolated copy, and stops at an approval gate. Nothing reaches your repository until you've seen the diff.
npm start -- Ada --shout
- Supported languages: english and spanish.
+ Supported languages: English and Spanish.
## Tests
Claude changed 1 file. Approve to commit, push and open a pull request.
Approve and open PRReject
// a run, start to finish
Five steps, and you hold the fourth.
A run is one piece of work: a prompt, a Claude configuration and a subject (a project, a task list or a single task). Schedules start runs on a cron, or you start one by hand.
Every step is recorded. The run page shows the live event stream, the full transcript, the diff, Claude's own summary and what it cost.
-
The schedule fires
Cron in your time zone, with retries, a time limit and a rule for runs that overlap. Or press Run now.
-
qubichelm prepares a workspace
A fresh copy of your GitHub repository or uploaded files, on the branch the schedule names. Your project, tasks and notes go into the prompt.
-
Claude does the work
In the mode you picked: answering, editing files, or running commands in a throwaway sandbox.
-
You review the diff
Approve, or reject and discard. Gates can also sit before a run starts, so nothing spends money without a yes.
-
qubichelm commits and opens the pull request
Signed off as qubichelm, with Claude's summary of what it changed and why in the PR body. Or commit and push only, or keep it local.
// pick how deep claude goes
Three modes. Each one reaches a little further.
You choose per configuration. Most work never needs a shell, so most work never gets one.
apianswers and reports
Read and report
One request to Claude with the relevant files and your project context attached. Good for summaries, reviews and status reports.
Changes nothing. No tools.
agentedits files
Work through the files
Claude reads, searches and edits the copy with qubichelm's own file tools, turn by turn, until the job is done. The default for code and docs changes.
Files in its copy of the repo only. No shell, no network.
cliruns commands
Run the tests too
The full Claude Code CLI, for work that has to execute something: install packages, run the test suite, build.
A sealed, throwaway container on a separate machine.
// where claude actually runs
When Claude needs a shell, it gets one far from anything that matters.
Command-running work goes down to a separate machine, into a container that exists for one run and is destroyed after it.
you
Your browser
Talks to the app and nothing else. You never reach the runner.
app.qubichelm.com
The app
Holds your keys in an encrypted vault and does every commit itself.
encrypted tunnel
The runner
A separate machine that talks only to the app, over a private tunnel.
one run, then gone
A throwaway container
Its own kernel, a read-only system, no privileges, and an allow-list for the internet.
- No real keys insideClaude's API calls go through a per-run proxy that swaps a one-time token for your key. The key never enters the container.
- Your git history stays putThe repository's history is read-only inside, and only the changed files come back. qubichelm commits from its own copy.
- Nothing left behindThe container and its files are deleted when the run ends, and a run nobody is watching is stopped.
// keep the books
Every run has a price and a paper trail.
What it cost, who approved it and what changed, written down every time.
Monthly AI budget
$41.20 spent$50.00 cap
A warning goes out at the percentage you choose. At the cap, runs stop until next month or until you raise it.
Cost per run
Input, output and cached tokens, priced per model, on every run.
Your own Anthropic key
Model usage bills to your Anthropic account. The key sits in the vault, encrypted.
Email when it matters
Runs that finish, fail or wait for your approval, and the budget warning.
An audit log
Who changed what, who approved which run, and which token did it.
// steer it from claude
Your backlog, reachable from Claude itself.
Point Claude Desktop or Claude Code at your organization's MCP endpoint with an API token. Claude can read projects and tasks, add work, check off to-dos and start runs, within the scopes the token allows.
The same tokens work against the REST API.
# projects and tasks list_projects get_project create_project list_tasks create_task update_task list_todos create_todo check_todo # runs trigger_run list_runs get_run list_schedules list_workspaces annotate
// start
Run your projects free. Add Claude when you're ready.
The free plan is a full project tracker: projects, task lists, tasks, to-dos and notes. Upgrade for Claude runs, schedules, the vault and your team, at $15 a seat each month.