Cockpit

My agent development control plane: every coding agent I use in one window, with the work around them made visible.

It runs locally on DeepSeek Harness. Claude Code, Codex, OpenCode and Hermes run as real terminals in the project I pick, each with its own login, skills and MCP servers. Around them sit an agent workflow, an image rail, automations, infrastructure and my knowledge vault.

I designed and built it with AI: ChatGPT for planning and images, Claude Code for building, Codex and DeepSeek for review. The film above is hand-written HTML, CSS and canvas, with a score synthesised in the browser.

kimlj.dev →

WorkbenchReal, persistent terminals for every coding agent, side by side in one project. Nothing about the agents is reimplemented.
Agent workflowOne prompt runs plan, build, test and an independent review. Findings go back for at most two automatic fixes, then validation, my approval and a checkpoint.
AgentsRoles are profiles with their own instructions and permissions, so any runtime can fill one, and a specialist is suggested when the task fits.
Image railClaude Code and Codex terminals get an image tool. Results land in the rail beside the conversation, and I approve or revise them in chat.
AutomationsJobs from Task Scheduler, systemd, cron, GitHub Actions and Hermes in one list, with instrumented runs reporting their real steps and retries.
InfrastructureMachines, containers, services and endpoints, read from the systems that own them with read-only checks.
KnowledgeMy Obsidian vault as a live graph. Agents read it before they plan and write back what they learned.