// code agents · on-premises
From a natural-language requirement to an approved merge request: an AI analyst turns your goals into tickets, agents write the code, and every change goes through peer-review cycles — with human control where it matters. Everything runs on your infrastructure: your code never leaves.
// how it works
This isn't autocomplete or a chatbot: it's a team of agents that runs your board end to end, under the same rules you'd expect from a developer.
Describe the goal in natural language. The analyst turns it into epics, tickets and milestones on your GitHub repositories — connected from the UI, with no manual token handling.
Your team prioritizes and assigns tickets to workers, sprint-planning style. Each worker operates in its own isolated git workspace, and as many run in parallel as your team can manage.
Every change goes through automated review rounds until it comes out clean. The PR arrives with context, numbered by round, for your final approval — nothing gets merged without human sign-off.
// what it feels like
mkanban follows the natural conventions of software engineering: backlog, planning, priority-based assignment and peer-review cycles. No agent does whatever it wants — there are AI and human controls at every stage.
Turns natural-language requirements into well-defined epics, tickets and milestones, ready to assign.
Keeps control: prioritizes the backlog, assigns tickets to each worker and gives the final approval on every merge.
Each worker can run a different model — stronger models for complex tasks, lightweight ones for simple work — and its own GitHub identity. LLM costs optimize themselves and history stays traceable.
Every change goes through automated review rounds before reaching a human — the same cross-checks you'd demand from a real team.
// what's included
Everything around the development cycle is built in — no fragile integrations.
Automated review rounds with a configurable limit. Infrastructure failures don't burn rounds: retries with backoff and model fallback.
The analyst turns goals into GitHub epics, tickets and milestones — a single source of truth for what's done.
File explorer, search, diffs and direct editing in the browser. Review and fix without leaving the board.
The state of every workspace and branch in one panel: what's running, what failed, what's waiting for review. Prometheus metrics included.
Unattended daily updates: automatic backup, healthcheck, and rollback only if something doesn't come up healthy.
Each worker can operate with its own GitHub account: every commit, PR and review comment is attributed to whoever made it. Full traceability in the history, like any other member of the team.
// true on-premises
mkanban installs on your server — a VPS, an EC2 instance inside your VPC, or a physical server in your own datacenter. Repositories, database and credentials live in your volumes.
// pricing
No seat licenses, no flat fees: mkanban charges 10% of the savings it generates, calculated on the tickets it actually closed during the month. A month with no output is a month with no invoice.
Calculated with the fixed parameters (4 h × USD 30 × 10%): equal to 10% of your real savings.
Equivalent to the output of 2 full-time developers.
mkanban runs on your server. You pay only 10% of the month's savings — no minimums, no fixed costs.
We operate your dedicated instance. Same 10% of the savings, with a monthly minimum that covers the infrastructure.
// faq
No. Repositories are cloned and worked on inside your server. The only outbound connections are the agents' calls to the AI model (with your API key, straight to the provider), GitHub, and a licensing heartbeat that reports usage counters — never code or content.
An x86_64 Linux server with Docker and HTTPS egress to four domains (we provide the exact list for your network team). Nothing else: everything ships in a single container.
Automatically, every night: the updater downloads the new version from the stable channel, backs up your data, brings it up and verifies health. If anything is off, it rolls back to the previous version automatically and notifies.
In arrears: tickets closed during the month × 4 h × USD 30 × 10%. Hours per ticket and hourly cost are fixed, deliberately conservative parameters of the model: if your real numbers are higher, your real savings are bigger and the invoice stays the same. Closed tickets come from your own instance's report. No black boxes: the number you see on the board is the number that gets billed.
Book a 30-minute demo on your own backlog. Even if you decide not to move forward, you'll walk away with a diagnostic of your development process.
Book a demo