> qubichelm

// how it works

A project tracker that can do some of the work itself.

qubichelm starts as a plain place to track projects. Connect a repository and an Anthropic key, and Claude can pick up the work on a schedule, with you approving what it changes.

// projects and tasks

The backlog is the brief.

Projects hold task lists, task lists hold tasks, tasks hold to-dos and notes. When a run starts, the project, its tasks and their notes go into Claude's prompt, so the work you already wrote down is the instruction.

  • Run against a whole project, one task list, or a single task.
  • Claude leaves a summary on every run: what it did, how sure it is, what to check.
  • Link related projects so a run can see across them.
greeter / docs
task
Fix the README's claims against the code
note
Language names are proper nouns.
to-do
Check the clone URL matches the remote
to-do
Check the documented output matches npm start

// schedules

Set the watch, then leave it.

A schedule pairs a prompt with a Claude configuration and a subject, and fires on a cron in your time zone.

  • Retries with backoff, and a hard time limit per run.
  • Decide what happens when runs overlap: wait, skip or run side by side.
  • Pick the branch: work on the base branch, an existing branch, or a fresh one per run.
  • Templates for the branch name, commit message, PR title and PR body.
  • Run now, any time, even when the schedule is paused.
schedule
name
Tidy README
cron
0 6 * * 1-5, America/Boise
config
default/agent/sonnet
branch
new branch per run: helm/{{run.shortId}}
commit
commit, push and open a PR
approval
required before commit

// workspaces and configurations

Tell it where to work and how much it may do.

A workspace is the code a run works on: a GitHub repository, reached with a fine-grained token you keep in the vault, or files you upload. A Claude configuration is the model, the system prompt, the mode and the limits.

  • Three modes: api to read and report, agent to edit files, cli to run commands in a sandbox.
  • Limits on turns and reply length; allow or forbid individual tools.
  • A smoke test checks a configuration before you trust a schedule to it.
claude configuration
mode
agent: edits files, no shell
model
claude-sonnet-4-6
max turns
40
tools
list, read, search, write, edit, delete
key
anthropic/production (vault)

// approval gates

Two places to say no.

Before a run starts, so a scheduled job can't spend money without a yes. And before anything is committed, so you read the diff before it reaches your repository. Set them for the whole organization, or per schedule.

  • The diff, the transcript and Claude's summary are on one page.
  • Approve and qubichelm commits, pushes and opens the pull request.
  • Reject and the work is discarded. Nothing was pushed.
run kv58aa5b

Waiting for your approval. The agent finished and changed 1 file. Review the diff, then approve to commit and open a pull request, or reject to discard the changes.

Approve and commitReject

// the run record

Everything a run did, kept.

Each run has its own page and its own folder of artifacts, kept for as long as you set.

  • Live events while it runs: each tool call, each result, each turn's cost.
  • The full transcript, the diff and the changed files.
  • Claude's summary, with its confidence and anything you should check.
  • Tokens and cost, rolled up into the monthly budget.
events
claude  Read /work/repo/README.md
result  32 lines
claude  Bash ls /work/repo
claude  Edit README.md
status  turn 4, 1,830 in / 214 out, $0.02
status  awaiting approval, pre_commit, 1 file

// your team

Organizations, roles, and one login across them.

Each organization's data is walled off from every other's in the database itself. Invite people by email; they can belong to more than one organization and switch between them.

  • Four roles: owners and admins run the organization (people, secrets, billing, gates), users do the work, viewers can look.
  • API tokens with scopes, for scripts and for Claude's MCP connection.
  • Email for finished runs, failures, approvals waiting and the budget warning.
api token scopes
read:all          read projects, runs, schedules
write:projects    create and update work
write:schedules   create and change schedules
trigger:runs      start runs
admin:secrets     manage the vault
admin:users       manage people
Start freeSee pricing