AI for software development teams works best around the code, not in place of the people who ship it: triaging bugs, summarising pull requests, keeping tasks and timers current, and flagging release risk. AlianHub gives engineering teams role agents for exactly that, plus Claude Code over MCP, while merging code and deploying to production are actions no agent can take.
Key takeaways
- Put AI on the coordination work around code: triage, summaries, status and release notes.
- Agents can move a task to In review. Only a person can call it Done.
- Merge and deploy do not exist as agent actions, so even a stolen token cannot reach them.
- Claude Code can work in your tracker as you, scoped to the projects you choose.
- CI keeps running where it runs today. AlianHub reads results, it never builds or deploys.
Where engineering time actually leaks
Developers rarely complain about writing code. They complain about everything around it:
- New bug reports with no repro steps, sitting unsorted
- Pull requests waiting for someone to explain what changed
- Tickets that say "In progress" for a feature merged last week
- Release checklists rebuilt from memory every sprint
- Status updates for people who will never open the board
Each item is small. Together they eat the focus time that good engineering needs. These are also the tasks where an AI agent can help without touching your codebase.
The AlianHub AI features engineering teams use
The Engineering and QA team pack
AlianHub ships role agents for 15 teams. The Engineering and QA pack includes:
| Agent | Job |
|---|---|
| Bug Triager | Sorts new bug reports |
| Code Reviewer | Supports review work on linked pull requests |
| QA Engineer | Supports QA tasks and checks |
| Release Manager | Helps keep release work on track |
Switching the pack on adds these agents to your projects, paused until you say go. You can undo the whole pack. See every pack on the Teams page.
Skills built for code teams
Skills are versioned playbooks with a dry run. Useful ones for developers:
- Summarise a pull request (pr.summary): reads the pull request a task links to, summarises the change and flags risk. It reads outside your workspace, so it stays off until the person who runs the server turns external reads on.
- Check a page for defects (qa-review): checks the page a task points to for defects. It needs a public URL in the task.
- Draft a PRD from a brief (prd.draft): drafts a product requirements document as a page that stays a draft until approved.
- Intake (brief.parse): proposes subtasks with estimates and the open questions.
GitHub and GitLab on every task
Connect GitHub from the Marketplace, pick repositories per project and see pull requests linked to tasks. GitLab links merge requests and commits to tasks. Sprints show progress with a burndown and a Gantt chart. Details are on the Integrations page.
The five-stage flow, with a hard stop
AlianHub's AI controls follow the same path for every task:
- Task: a person writes the brief and assigns it.
- Work: the agent keeps the task and timer current. It may set In progress or In review, never Done.
- Result: it posts what it found and attaches the work, such as a link to the pull request.
- Approval: it proposes only, for example asking to publish to the test site. A person decides in the AI Inbox.
- Hard stop: deploy to production is human only. Attempts are logged.
Your CI is still the source of truth. AlianHub reads the result from GitHub or GitLab and never builds or deploys. Read the full model on the AI controls page.
Tip: give the Bug Triager L2 so it can comment and change status within a list on its own, but keep moving tasks between lists and assigning work in the AI Inbox until you trust its judgement.
Claude Code and other CLI agents
Many developers already live in Claude Code. AlianHub ships an MCP server, so Claude Code can read and update your tasks while you work. Claude Code, Codex and other CLI agents use the Agent API from a personal account, scoped to chosen projects, and their writes land in the AI Inbox first.
There are three ways to run agents, chosen by owners and admins:
- Workspace key: one API key in settings, paid by the company and capped per agent. Best for scheduled agents.
- Personal AI account: a developer's own Claude Code, on their machine, with no key held by the workspace. Revoke the token and the link dies.
- Self-hosted model: Ollama or vLLM on your GPU, inside your network. Best for regulated teams and air-gapped installs.
How to do this in AlianHub
- Open Settings > Import & export and bring your backlog from Jira, ClickUp, Trello, Asana, Monday.com or CSV, or start a new project.
- Connect GitHub or GitLab from the Marketplace and pick the repositories for the project.
- Switch on the Engineering and QA team pack. The agents join paused.
- Open AI > Agents, choose the Bug Triager and set its autonomy level. Set a daily action limit and monthly spend cap.
- In the Skill library, run a Dry run of a skill on a real task to preview what it would write.
- On AI > Connect your AI, choose Create a token, pick the projects and permissions, and add it to Claude Code. The token is shown once.
- Review proposals in the AI Inbox and every change in AI > Audit log.
A business scenario: a small product team ships more often
A product studio with a handful of developers runs two-week sprints. Bug reports arrive by email and chat, and the tech lead triages them by hand before standup. Pull requests wait while reviewers work out what changed.
With the pack on, new bugs are sorted and commented before standup. The Code Reviewer's summaries give reviewers context first, and Claude Code updates task status as developers work, so the board matches reality. The tech lead gets back hours of coordination each sprint, reviews move faster, and the team can take on more client work without hiring. Merges and deploys still go through the same people and the same CI as before.
What agents never do
For code teams, these limits matter most. No agent can:
- Merge code or deploy to production
- Delete a project or a task
- Remove a member or change permissions
- Mark a task Done
These are not settings you might forget. They do not exist as actions. Everything else an agent does is written to the audit log, and changes it makes on its own come with Undo.
If you are coming from Jira, our post on the Jira alternative for small teams covers the move.
FAQ
How can AI help software development teams without touching the code?
Use it for coordination: triaging bugs, summarising pull requests, keeping task status and timers current, and drafting PRDs and release notes. In AlianHub, merging and deploying are not available to agents at all.
Can Claude Code update tasks in AlianHub?
Yes. Create a token on the Connect your AI page, choose the projects and permissions, and add it to Claude Code. It acts as you, sees only what you can open, and its changes are logged.
Can we use a self-hosted model instead of a cloud AI?
Yes. AlianHub supports a self-hosted model through Ollama or vLLM on your own GPU, so nothing leaves your network. It is slower on hard reasoning, but fits regulated and air-gapped teams.
Will an agent close tickets automatically?
No. No agent closes a task in any project. A person checks before Done, whatever the project has chosen.
Try the Engineering and QA agents in the live demo, self-host AlianHub free, and share this with your tech lead.