The shared working layer, in detail

The developer and the AI agent work on the same project objects

This is the page for anyone who wants to understand why SpecViber isn't just a nicer surface, but a real shared working layer for AI-assisted development.

Each side through the surface that fits it best. The developer in the Web UI and VS Code. The agent through MCP tools, instructions, and skills. Same documents, same task objects, same state, same routines.

Request an invite Contact us
Project layer Intelligent markdown Tasks & action specs Kanban Graph Chat Shared context Web UI + VS Code Verification
Under the surface

One shared store underneath

Document metadata, checklist and comment states, relations, assignments, history, task information — readable and writable from both sides.

Your markdown and your code stay in your project; SpecViber holds the metadata that coordinates the work.

Shared IDs Explicit relations Close-guard before done
The shape of the layer

Developer surfaces on one side, agent surfaces on the other, one shared store in the middle

The point is not that both sides see the same UI. The point is that both sides read and write the same coordination layer: document metadata, checklist states, relations, assignments, history, and task information.

That is why the agent stops working from a loose snapshot and starts working against the same live project state as the developer.

Diagram showing developer surfaces and agent surfaces connected through one shared metadata store
Developer surfaces and agent surfaces meet at one shared metadata store.

Built for your environment, not as a replacement

SpecViber drops into an existing setup because it doesn't replace the foundation; it adds a working layer around it. Existing repos, existing markdown, existing IDE, existing LLM, existing habits. Because the agent side is MCP-native, it works with the assistant you already use — Copilot, Claude Code, Cursor and others — with no vendor lock-in.

Your code and your documents never leave your machine. SpecViber stores only the metadata it needs to coordinate the work — states, relations, assignments, history.

Markdown that does the work

SpecViber starts from ordinary markdown because that's already where developers write specs, notes, design decisions, READMEs and implementation plans. It doesn't lock those files into a proprietary format; it makes them operational.

A markdown file becomes a structured work object with status, progress, comments, relations, checkboxes, metadata and task connections — useful for the human reading it, the agent acting on it, and the system tracking the state of the project. The document stops being passive text and becomes part of the workflow.

Properties

Clear, readable metadata close to the content.

Inline comments

Discussion in the right place in the right document.

Tracked checkboxes

Real work states — todo, in progress, done, dropped, blocked, needs research — readable while work happens.

Relations

Documents and work objects explicitly linked: references, depends_on, part_of, supersedes.

Status and progress

The state of the work becomes visible without the document ceasing to be readable.

Action spec at the centre, task for flow

A task is a work object that's easy to pick up, assign, and track through a flow. An action spec is a living markdown document that carries the work from the first problem statement to a checked, documented delivery — problem, plan, steps, sub-specs, progress, and verification in the same structure.

Task for flow. Action spec for carrying work end-to-end with the agent and you working against the same plan.

Kanban as one live cockpit

Tasks and action specs share one Kanban flow, so the board is more than a planning view — it becomes a live cockpit for developer-led AI work. See what's waiting, in progress, blocked, or ready for review. Assign work to the agent, follow the card, open the related spec, and watch progress update as the work moves.

Because the board, documents, comments, tasks, relations, and verification steps all point back to the same shared working layer, the whole project becomes easier to steer.

Relations you can see and walk

As a project grows, search isn't enough. You need to understand how things connect. SpecViber makes relations explicit: a document can reference another document, a spec can depend on another spec, a decision can supersede an older one, a task can belong to a larger action spec, and one direction can conflict with another.

Those relations aren't just metadata buried in a table. The relation viewer lets you explore them graphically: from any object — document or task — you can see related objects and navigate straight to them, with key information surfaced as you go: progress, status, assignments, content and more.

That helps the developer navigate the project, and it helps the agent load the right context before acting. The project becomes not just searchable, but understandable as a connected system.

Chat as a driving surface, not the source of truth

Chat is often the fastest way into work — that's where you ask the agent to investigate, break down, write, pick up, and suggest next steps. The difference in SpecViber is that the agent doesn't work in a vacuum. It works against the same documents, tasks, relations, and rules as the rest of the system, and the state of the work stays readable in the project even after the chat scrolls away.

Shared context for developer and agent

Developer and agent share the same documents, the same task objects, the same relations, the same IDs, the same rules, and the same routines. The agent stops working from a loose snapshot of the project and starts working against the same live working layer as the developer — which is what makes the collaboration stable over time.

Web UI and VS Code extension

The web UI is your overview and dashboard surface. The VS Code extension is your document-near collaboration surface, close to the files. Both should be described as parts of the same method, not as competing interfaces.

Web UI overview

Web UI — overview and dashboard

VS Code extension

VS Code — document-near work

Done because it's verified, not because the chat says so

AI-assisted work often fails at the last step. Something looks done, the chat says it's done, but nothing was actually checked. Every action spec can end with explicit verification steps. The agent runs the checks it can handle itself — terminal commands, syntax checks, tests, API calls, other automated validation. Anything needing human judgment, UI interaction, or external confirmation stays visible for the developer.

A close-guard keeps a spec from being marked done while open steps, failed checks, or unmet criteria remain — so a spec isn't complete because the agent says so, but because the plan, the progress, the checklist, and the verification all point to the same conclusion. Every step is logged and timestamped, so closing a spec leaves a record rather than a vague “done.”

That's how SpecViber keeps vibe coding honest.

Close-guard flow showing open steps, automated checks, user review, and verified done

The model fits — what's next?

If the mechanics resonate with the way you want to work, request an invite or read how access works.

Request an invite Read Pricing & access

FAQ

What does “markdown that does the work” mean in practice?

Regular .md files that gain status, progress, comments, relations, and tracked checkboxes — without ceasing to be regular, readable markdown files in your repo.

Why aren't plain tasks enough?

Because some work needs to carry plan, execution, and verification in the same object. That's what action specs do — and why they sit at the centre of the workflow.

Why is verification its own part of the model?

Because AI-assisted work often fails at the last step — the chat says it's done before anything was actually checked. Close-guard plus explicit verification steps keeps a spec from closing while open work remains.

Does this lock me into a specific AI assistant?

No. The agent side is MCP-native, so it works with the assistant you already use — GitHub Copilot, Claude Code, Cursor and others.