I have seen demos where an agent picks up a tagged task, makes a pull request, and adds the link back to the task. What would I actually need to run something like that inside my network? And where do MCP and command-line tools fit in?
How would I host an agent that turns kanban tasks into pull requests?
Answers
You can start with a pretty ordinary service. Have the kanban tool send a webhook when a task gets the right tag, then start a worker for that task. The worker reads the task, checks out the repository, asks a model to help make the changes, runs the available checks, and opens a pull request. Finally, it posts the link back to the task.
The agent part is the loop between the model and the worker. The model can request an action, the worker performs it and sends the result back, and the model decides what to do next. You can host the worker in a container without adopting a whole agent platform. The model might still be an external API unless you also host one yourself.
Reply to this answer
MCP is a protocol for exposing tools and context to compatible clients. It is not required for an agent, and it is not the same thing as the model or the worker that runs it.
For one workflow, calling a CLI or your own functions can be simpler. An MCP server may be useful if several clients need the same integrations. Hosting one just means running a service or process that exposes those capabilities. Any pricing depends on the service or tools behind it, not on MCP being inherently metered.
Reply to this answer
Keep the code execution isolated. An agent working on a repository may run tests, install dependencies, and execute commands, so a dedicated VM or a suitably restricted remote environment is a safer starting point than your everyday machine.
Give it narrowly scoped credentials for the task and repository, and limit network access where practical. Treat the pull request as a proposal for human review, not as permission to merge or deploy automatically.
Reply to this answer
Before building the always-on part, try a coding agent on a small, well-defined task. That gives you a feel for what needs human intervention and what checks you want before spending time wiring up the board and repository integrations.
Reply to this answer
Add your answer
Answers are reviewed before publication.
2k~canvasgrommet_539:
What does giving the model a tool mean in that setup?
_f29!mivqueno.7:
You describe an available function and its arguments, such as reading a task or opening a pull request. The model requests a call, but your worker executes it and returns the result.