Knowledge base and dashboard powered by agentic development

We built our own knowledge base (clients, projects, tasks, contacts, links) as git-versioned markdown files and an interactive dashboard on top of it that does more than display content: it edits it, runs AI agents on tasks, syncs both ways with ClickUp, estimates time and cost, and visualizes how projects relate. The whole system is built and extended through agentic development — AI agents ship features in small, verified steps, with automatic git versioning.

What you can do with it

Knowledge base

• Clients and projects overview — Client → Project → files (plan.md, history.md, specification.md, contacts, links).
• In-browser editing (markdown editor) with a git commit on every action.
• Automatic commits — safe by design: never pushes on its own and never commits .env.

Tasks

• Tasks as markdown checklists with metadata — complete them and sort by status and due date.
• Clear distinction of who did the work — human vs. agent.
• AI estimates on every task: human time (h) and agent cost (€ in tokens), with milestone and project totals and a human-vs-agent comparison.

Search and connections

• Global search (Cmd+K) across clients, projects, files, and individual tasks.
• Knowledge graph — semantic similarity of tasks and projects via embeddings (not just keyword match), cross-links, and a configurable similarity threshold.

Agents

• Launch AI agents on specific tasks directly from the plan.
• Agent monitoring — what they are doing, which model is running, how long, live logs; stop / rerun.

Dashboards

• Home dashboard: stat cards, throughput chart, project progress, human vs. agent, tracked time, and a period filter.
• Charts as plain SVG — no heavy charting libraries.

What we connected

• Git — versions the whole base, commits on every action, auto-commit.
• ClickUp (REST API v2) — one-click two-way task sync (status, assignee, due date), import of new tasks, and tracked time per person and project.
• Anthropic API (Claude) — powers AI agents and task estimates.
• OpenAI API — embeddings for the semantic knowledge graph and estimates, with a lexical fallback.
• Claude Code CLI — headless agents that actually do the work.

AI models we use

Claude (Opus, Sonnet, Haiku) powers agents and estimates. OpenAI (GPT and embeddings) handles estimates and semantic similarity in the graph. Model pricing (input/output $ per 1M tokens) is built in, so token estimates convert into a real euro cost per agent task.

How we build it — agentic development

This is not classic hand-written coding. AI agents (Claude Code) write and change the code; a human sets the goal and approves the result.

• We proceed in small, verified steps — every change is type-checked and verified live in the browser (screenshot / interaction) before it is committed.
• Multi-agent workflows — on larger tasks, agents run in parallel on separate parts, then we integrate and verify.
• The dashboard itself can dispatch headless Claude agents onto its own tasks — so the tool both uses agentic development and enables it further.

Impact

• One place for clients, projects, tasks, estimates, and metrics — instead of scattered documents.
• ClickUp stays the source of truth for status and time, while we plan and analyse in one place.
• Measurable work economics — on every task we see human time and agent cost, and can decide which approach pays off.
• Fast iteration — new features land agentically, continuously, and safely (versioned in git).

Looking for a similar internal tool, knowledge base, or agentic workflow on top of your data? We will gladly design a custom solution.

Get in touch