Tool · Build
The Workbench is where developers build with coding agents on real projects, on your own infrastructure. It turns a ticket into a plan, runs agents in isolated workspaces, and keeps a person in control of every result.
Capabilities
Turn a ticket or an intent into an implementation plan and break it down into concrete subtasks before any code is written.
Run several coding agents in defined workflows, each handling part of the work, instead of one agent doing everything at once.
Works with OpenCode, Claude Code, Aider and CLI-based agents, so teams are not locked into a single agent runtime.
Work across several repositories at known, pinned commits, so a task is well defined and the result is reproducible.
Each agent works inside an isolated, sandboxed workspace rather than loose on the host, so it only touches what it is meant to.
A person stays in control. Agents can propose and mark work done, but the human keeps the override and the review stays binding.
Designed to sit alongside your existing Bitbucket and GitLab workflow, with an audit trail of every agent action.
Use the model you selected in the Benchmark. Switch models without rewriting endpoints, on self-hosted or cloud compute.
Security · Harness Security
A coding agent is more than a model. Around it sits the harness: the instructions it starts with, the tools it may call, the rules that approve or block each call, the context it loads and the agents it hands work to. Whoever controls the harness controls the agent – so that is where we put security. Rights are granted per task step and enforced by deterministic checks the model cannot argue with. Everything the agent reads – tickets, READMEs, dependency docs – is treated as data, not orders. And when several agents work together, each gets its own workspace, its own rights and an independent reviewer.
01
Whoever controls the harness controls the agent.
02
Your agent reads everything. So does your attacker’s text.
03
More agents, more doors. Each one gets its own key.
Security · Gates, identity and review
Agents produce code faster than any security team can read it. So the checks move into the loop: every agent step runs through the gates your pipeline would apply later, and the agent fixes findings before a person sees the diff. Every change carries the identity of the agent, the model and the person behind it. And because a human in the loop only helps if the human can still see, we design review for real attention: small diffs, risk-based routing and reviewers who know what AI-written code gets wrong.
01
Every change is checked. Every change has a name.
02
A human in the loop only helps if the human can still see.
Workbench · Positioning
Not a speed tool for vibe coders, but the controlled entry into agentic development — especially with on-premise models.
01 — Beginners
Training wheels for getting started with AI — especially with on-premise models, where context and steering are more demanding. It tends to slow the hardcore vibe coder down.
02 — Enterprises
Good software engineering — traceability, tests before implementation, quality gates, templates — becomes part of the AI process instead of the wild west.
03 — Compliance
Full audit trail, managed context instead of uncontrolled on-the-fly loading of skills or plugins — AI hacking is contained, teams work in a standardized way.
How it looks
Ticket
Start from a ticket or an idea.
Plan
Plan and decompose it into subtasks.
Agents
Agents work in isolated workspaces.
Review
Review the result, with a full audit trail.
Handover
Hand over to your normal review and merge.
What it gives you
Traceability & quality, built in
The Workbench bakes testing, quality assurance and traceability into the development process itself. An optimised agentic workflow produces results on your own on-premises models — tested as they are built, and traceable through every agent step back to the original requirements and specifications.
The approach
Runs fully offline inside your own data centre. No data egress, no external API calls, no hidden telemetry channel.
European LLMs on your own hardware. Updates via a controlled supply chain: you decide what is rolled out, and when.
A full audit trail of every agent action. RBAC, SSO and SIEM integration address NIS-2, ISO 27001, GDPR and BaFin requirements.
