Case study: AI agents and automation
Open-Source 24/7 AI Agent with Human Approvals
Runs scheduled work around the clock, and nothing leaves the machine without a person's approval.
A self-hosted agent that drives Claude Code from Slack and cron, with sealed approval cards, exactly-once scheduling and no secrets in the model's reach.
- Where
- Open source (MIT) on GitHub
- Role
- Author and maintainer
- When
- 2026
- Team
- Solo, end to end
- 1
- run per cron slot, claimed in SQLite and safe across restarts
- 240
- Vitest tests over scheduler, approvals and security paths
- 0
- secrets in the model's environment; root actions only on approval
- TypeScript
- Node.js
- Claude Code
- MCP
- Slack Bolt
- SQLite
- systemd
- Vitest
TL;DR: tara-agent is an open-source, self-hosted AI agent that runs around the clock. It drives the official Claude Code CLI from Slack threads and cron jobs, and it stops for a human whenever an action matters: sending a message, a risky change, a login. Approvals are cards in Slack or on a web dashboard, every scheduled run happens exactly once, and the model never sees a secret. Source: github.com/abhijeet-maharana/tara-agent.
What problem does it solve?
AI coding agents are good at real project work, but most setups only run while someone is watching a terminal. I wanted an agent that keeps working on its own and still asks before anything leaves the machine:
- Work happens on a schedule, day and night: reports, checks, drafts, fixes.
- A person decides before anything is sent, published or changed outside the repo, usually from a phone.
- A bad prompt or a compromised page can't reach the keys, because the agent runs with the least access possible.
How does it work?
One Node.js process per host. It never calls a model API directly: every piece of model work is a headless run of the claude CLI inside the right project repo, so each project keeps its own instructions, skills and MCP servers.
- Slack in, Claude out. A message in a mapped channel starts a Claude Code session; replies in the thread resume it. Progress streams live into the thread as a task list, then the answer.
- Scheduled jobs come from a
jobs.jsonin each repo, hot reloaded. A cheap pre-run script can decide there is nothing to do and skip waking the agent, which keeps model usage low. - A built-in MCP server gives Claude its tools for people:
request_approval,propose_item,send_fileandrequest_privileged. Each run gets its own token for it. - SQLite stores sessions, runs, slot claims and approvals, so a restart loses nothing.
How does it guarantee a job runs exactly once?
- Slot claims. When a cron slot fires, the scheduler claims it in SQLite first. A claimed slot never runs again, even after a crash or restart.
- Catch-up. On start, each job runs its most recent missed slot once, if it is recent enough.
- Named locks keep two jobs that share a resource (for example one browser) from running at the same time.
- Interrupted runs cut short by a restart run again once on the next start.
How do human approvals work?
- Cards, not chat. Drafts wait as cards that can be edited, approved, rejected or sent back for a new version. A decision continues the same Claude session, as the same job.
- Queues send approved items out later in delivery windows per region, with daily and weekly caps, gaps and batches.
- Sealed approvals. Each approved item carries an HMAC of its kind, id and approved field values. An executor in another process verifies the seal before it acts, so editing a file by hand never sends anything.
- Project checks. A project can attach its own blocking checks to a card before it can be approved.
How is it secured?
The threat model starts from one rule: anyone who can post in a mapped channel can make Claude act on the machine.
- Unprivileged user. It runs as its own OS user with access only to the project repos, under systemd.
- Split users. Claude and the pre-run scripts can run as a second OS user, so a run can't read the agent's database or keys.
- No secrets in the model's environment. Slack tokens and other credentials are stripped before the CLI starts. The system prompt is passed in a file so sudo never logs it.
- Root actions on approval. The agent never gets sudo. It can request an action from the allowlist of a root-owned helper, and a person approves it.
- Dashboard decisions need an unlock cookie set from another device.
What else did I build into it?
- A web dashboard, server-rendered and phone friendly: what's running, what needs a person, failing jobs, the day's schedule with overlaps and free time, and full transcripts of every run.
- A read-only Code page that shows each repo's uncommitted changes and diffs through a private copy of the git index, so it never touches the real repo.
- Learning from mistakes. After a run that had trouble, the agent gets one chance to fix the instructions it followed.
- A daily health digest per project: runs, problems, missed slots and peak usage.
What were the results?
| Metric | Result |
|---|---|
| Status | In daily use, running scheduled jobs for my own projects |
| Codebase | About 10,000 lines of TypeScript, MIT licensed |
| Tests | 240 Vitest tests over the scheduler, approvals, sessions and security paths |
| Dependencies | 5 at runtime: MCP SDK, Slack Bolt, croner, yaml, zod |