Three layers, one purpose: make AI agents auditable teammates.
Fabriq sits between your agent and your repository, intercepting every capability call and surfacing it for human review.
Desktop Shell
dapei-engine
Agent Backend
Electron + React — two-dimensional UI with confirmation-driven surface
Scanner → Validator → Commit pipeline — 100% deterministic, no LLM calls
Stdio JSON-RPC — works with OpenCode, Claude Code, or any ACP agent
Three surfaces that cover the full AI development workflow.
From browsing your workspace to closing a feature — every stage has a dedicated UI surface.
Workspace Browser
A two-dimensional surface that separates long-lived workspace assets from short-lived feature worktrees. The shell enforces the boundary — the agent can never cross-write.
- Dual-dimension routing (workspace vs feature)
- Long-lived asset management (specs, contracts, tokens)
- Short-lived feature worktree isolation
- Visual boundary enforcement in UI
Feature Workbench
A stage-driven cockpit that takes a feature from spec to shipped code. Each stage produces structured artifacts with evidence — not just code.
- Stage-gated workflow (design → implement → review → close)
- Structured artifact generation per stage
- Confirmation prompts before state changes
- Context injection into agent sessions
Scanner + Validator Engine
LLMs extract cognitive artifacts from your codebase; the deterministic engine revalidates every claim against source before committing — no trust required.
- LLM-powered extraction (scanner role)
- Deterministic revalidation (validator role)
- Source mismatch = artifact rejected
- Full audit trail of every decision
How it works in practice.
Three steps, from connecting your workspace to shipping with confidence.
Connect your workspace
Point Fabriq at a directory containing your repos, specs, and feature worktrees. Attach OpenCode or Claude Code as the agent backend. Fabriq discovers the workspace structure and maps the two dimensions automatically.
Agent proposes, shell surfaces
Every capability call from the agent is typed, dimension-aware, and routed through a validated IPC layer. Destructive actions don't execute silently — they queue in the UI as explicit prompts showing the target, payload, and expected effect.
You confirm, then ship
Review each action, approve or reject, and keep a full audit trail. Every decision leaves sourced, validated evidence. Reversal is a first-class operation — not a post-mortem. No more guessing what the agent just did.
Core design principles.
Every feature in Fabriq traces back to one of these five concepts.
Typed Capability Channel
Every state change goes through a validated IPC layer with schema-checked payloads. If a capability isn't in the allowlist, it doesn't execute. Prompt injections can't mutate your disk.
Two-Dimensional Workspace
Workspace assets (long-lived specs, contracts, tokens) and feature worktrees (short-lived agent branches) are separate dimensions — at the routing level, the IPC layer, and in the UI.
Confirmation Gates
Destructive capabilities surface as explicit prompts with full payload visibility. You review the target, see the intent, and approve or reject. Nothing hits disk until you say so.
Scanner + Validator
LLMs extract cognitive artifacts from code slices; a deterministic engine revalidates every claim against the source before committing. If the artifact doesn't match the source, it doesn't get in.
Evidence-Based Artifacts
Every decision leaves provenance: facts carry source references, inferences carry derivations, unknowns carry reasons. Three confidence levels with mandatory evidence fields.
Agent-Agnostic Protocol
One ACP stdio JSON-RPC bridge works with OpenCode, Claude Code, or any ACP-compatible agent. No PTY shims. Share one agent process across multiple UI surfaces.
Technical specifications.
Open source, cross-platform, built with modern tooling.