FEATURES

An audit trail that survives the agent.

Three pieces — a single write path, a knowledge graph, and GitHub as the source of truth — that turn agent-driven engineering work into something you can review, redirect, and inherit.

Single write path through MCP

Agents propose plans and add tasks through `propose_plan` and `add_task` — there is no human form to create work. Humans approve, redirect, or reject. The result: a clean audit trail where every node was either machine-authored or human-approved.

  • Plans land as `proposed`; humans flip to `active`
  • 15+ tool schema validated against Zod at the edge
  • Per-tool OAuth scopes granted at connect time

Knowledge graph as long-lived memory

Knowledge entries with `supersedes` edges form a temporal graph. Search hybridises pgvector HNSW with keyword scoring, and freshness decay nudges old facts out. Agents learn from prior plans across the org without you copy-pasting between chats.

  • pgvector HNSW + BM25 hybrid scoring
  • `supersedes` edges make truth tellable from history
  • Per-org isolation; nothing leaks across tenants

GitHub is the ground truth

A task does not transition to `done` because an agent says so — it transitions because the PR merged, the expected files changed, and CI is green. Anything ambiguous lands in `needs_review` for a human gate.

  • GitHub App + GitLab + self-hosted GitLab supported
  • Per-repo `required_check_runs` + `verify_branches`
  • Failed verification → `needs_review`, never silent

SOVEREIGN BY DEFAULT

Built in EU. Data stays in EU.

Drobek is an EU-native product, not a US tool with a European checkbox. Our sub-processor list is short, GDPR is the primary framework, and the only US transfer is OpenAI for embeddings — under Zero Data Retention.

  • Hostinger EU VPS for primary storage; EU SMTP for transactional email
  • Sentry.io EU region for errors
  • OpenAI with Zero Data Retention enabled — knowledge bodies are processed for embeddings only
  • GDPR-primary Privacy Policy (not an adapter)
  • Privacy Policy + Cookies + Terms + DPA available in CS / EN / DE / ES

Read the privacy policySee pricing

HOW IT WORKS

Four minutes from sign-up to a running agent.

  1. Create a workspace. Slug auto-derived from the name.
  2. Drop the MCP config into Claude Code / Cursor / Claude Desktop. Copy button; one-click into the right config file. Your client signs in via OAuth — no tokens to copy.
  3. Ask your agent to propose a plan. It lands in /plans for your approval.
  4. Approve. Watch the agent claim work, open PRs, and verify against CI.

Start free