Standing company context
Company × role × project markdown is loaded into every interaction. No more re-explaining the business to your tools all day.
The Cursor of Prose.
Collaborate on files using multiple AI models.
MdCoWork is a company-wide AI workspace where every chat already knows your business, team and project context, and markdown files created by AI are editable and collaborative right in the chat window. Share files with models and colleagues; work together in realtime - everything is versioned by our intelligent Auto-Commit engine.
Positioning refresh
Sharpen our positioning line — make it lead with what we do.
Done. I pulled audience from strategy.md and revised positioning.md — review the doc →
Reply to MdCoWork…
1.2 KB · 12 lines · Syntax visible, human/AI editable
# Positioning
## One-liner
-We help teams use AI better.
+MdCoWork is the AI workspace
that already
+knows your business —
context-loaded, not cold.
## Audience
+- Teams of 20–500 already
running on AI
v14 · reversible
Save money on AI subscriptions
Every model, one workspace. Switch between the latest frontier models, or let auto mode pick per task.
Twin Mode: Two is better than one
Twin mode pairs two models together check each other's work and measurably raise answer quality.
Auto-Commit
Seamless, transparent git-like versioning without the hassle of remembering to save/commit, or writing a summary. Real time conflict resolution via AI with recovery points that finally make sense
- 1 brain
- Shared across the whole company
- 3 platforms
- Use in the browser, or download for Windows and macOS
- All frontier models
- Access all of the latest AI LLMs without individual subscriptions. Combine models for extra power!
The problem
Teams running on AI have no shared brain.
Teams running on AI have no shared brain. Every employee's agent starts cold — ignorant of company practice, role context, and what was decided last week — so people re-explain the business to their tools all day, and answers drift because everyone's context is different.
Generic chat assistants
Don't know your company. Every session starts from zero.
Google Docs / Notion
Know your documents, but aren't agent-native and can't act.
Enterprise RAG search
Read-only Q&A over your docs — not a place to write and edit.
The solution
An agent workspace pre-loaded with your org's knowledge.
- Org structure is the context index — company, role, and project markdown loaded every interaction. Curation, not retrieval.
- Agents produce editable markdown versions; permissions decide who can change what.
- Everything is versioned via mddbase — every agent write is attributed and reversible.
- The store stays dumb; the brain lives in MdCoWork.
A new category
Collaborative Markdown
Cursor changed how developers write code: the AI proposes, you dispose, line by line, inside the document itself. MdCoWork brings that exact loop to prose — your company’s knowledge, plans, and decisions, written together in plain markdown.
Cursor is for code
Agents and developers edit the same files, with inline diffs and instant review.
MdCoWork is for prose
Agents and teams edit the same markdown, with the same propose/dispose loop — for everyone, not just engineers.
- Markdown is the lingua franca of AI — readable by humans, native to models, diffable by design.
- Every document is a shared surface: humans type, agents edit, both see the same versioned truth.
- No silent overwrites. Collaboration, not automation — the human always disposes.
Collaborative Markdown, for the whole team
Whole-team logins
Shared API access to frontier models removes the per-seat cost of giving everyone an Anthropic or OpenAI account.
Propose, then dispose
Agents produce editable markdown versions; humans accept, edit, or reject. Permissions decide who can change what.
Everything versioned
Built on mddbase, every agent write is attributed and reversible — git-like history you can actually read.
Markdown that learns
"Those figures are out of date — update them" is a valid, actionable, versioned prompt, not a chore.
Collision-safe collaboration
Humans and agents edit the same files at once. Last-writer-wins, nothing destroyed, history always recoverable.
How it works
Agents propose. Humans dispose.
The consequential layer is always reversible. Nothing the AI does to your knowledge base happens without a reviewable, versioned step you control.
Context loads automatically
Company, role, and project markdown is included as standing context — every chat already knows who you are and how the business works.
The agent proposes
Ask it to draft, revise, or update. It returns an editable markdown version with a clear, reviewable diff — never a silent overwrite.
You dispose
Accept, edit, or reject. The accepted version is committed to mddbase, attributed and fully reversible.
Why combine models
A panel of models beats any model alone.
MdCoWork lets you fuse models — run several together and reconcile their answers, instead of betting everything on one. The result is consistent: every fusion panel outperforms its strongest member working solo. Twin mode is the simplest case (two role-different models checking each other); larger panels push quality higher still.
Notably, the best fusion of mid-tier models can match or beat a top model running alone — so you don't need the single most expensive model to get top-tier answers.
Fusion panel Solo model Scored on 93 of 100 tasks
Fable 5 + GPT-5.5 (fusion)
69.0%
Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro (fusion)
68.2%
Opus 4.8 + GPT-5.5 (fusion)
67.5%
Opus 4.8 + Opus 4.8 (self-fusion)
65.6%
Claude Fable 5 (solo)
65.4%
Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro (fusion)
64.6%
DeepSeek V4 Pro (solo)
60.4%
GPT-5.5 (solo)
60.0%
Claude Opus 4.8 (solo)
58.6%
Kimi K2.6 (solo)
53.8%
Gemini 3.1 Pro (solo)
45.4%
Gemini 3 Flash (solo)
43.0%
40%45%50%55%60%65%70%
Benchmark: 100-task internal evaluation. Axis truncated to 40–72% to show separation. Orange entries scored on 93 of 100 tasks.
Roadmap
Revenue and difficulty arrive together.
Phase 1 · Shippable now
Turn-based workspace
- — SSO and MLAuth paired identity for agents acting as or with users
- — Cowork-style chat with standing company + role + personal context
- — Reviewable, permission-gated edits via mddbase
- — Collision detection with recoverable history
Phase 2 · The premium moat
Real-time co-editing
- — Live, character-level co-editing — human typing while agents edit the same file
- — OT/CRDT engine in the session layer, checkpointed to mddbase
- — The literal Cursor-for-prose experience
Future
Knowledge that compounds
- — Per-file reputation — PageRank for the company knowledge base
- — Gap detection: the model reports what context was missing
- — Intelligent merge via twins — propose, never dispose
The md stack
Build order is also usability order: each layer is useful before the next exists. MLAuth gives agents an identity; mddbase gives them somewhere versioned to write; MdCoWork is where teams actually work. "MdCoWork on mddbase" tells the layered-stack story in two words.
Identity · free standard
LiveMLAuth
OAuth for agents. A public-key identity layer so an agent can prove who it is and which human authorised it — no passwords, no shared secrets.
Storage · infrastructure
In buildmddbase
Supabase, but for markdown. A dead-simple hosted backend that stores markdown in folders, versions every change, and stays deliberately dumb so anything can build on it.
Application · the product
In designMdCoWork
Where the brain lives. Per-seat SaaS for teams: the pre-loaded company context, the propose/dispose edit loop, twin mode, and real-time co-editing.
“Isolate the consequential from the advisory, and make the consequential reversible. Agents propose edits; humans dispose.”
The guiding principle, inherited across the whole stack.
Download
Every platform. One workspace.
Use MdCoWork in the browser, or install the native app on your phone or desktop. Same context, same markdown, same propose/dispose loop — wherever you work.
Desktop and mobile apps are in early access — buttons are placeholders for now.
Pricing
The value lives in the application layer.
MdCoWork is per-seat SaaS. The infrastructure beneath it (MLAuth, mddbase) stays cheap and neutral by design — you pay for the brain, not the plumbing.
Solo
Free to start
Try the workspace and the markdown edit loop on your own.
- Single user
- Bring your own API keys
- Versioned personal markdown
- Model switching
Team
From $20 per seat / month
For teams ready to give everyone a context-aware AI workspace.
- Whole-team logins, shared model access
- Standing company + role context
- Reviewable, versioned markdown edits
- Model switching + auto mode
Enterprise
Let's talk custom
For larger orgs needing SSO, granular permissions, and twin mode at scale.
- SSO + MLAuth paired identity
- Granular per-file permissions
- Twin mode + real-time co-editing
- Priority support & SLAs
Indicative pricing — not yet live. We are onboarding design partners first.
Give your whole company one brain.
We are onboarding a small group of design partners now. If your team already runs on AI and is tired of re-explaining the business every morning, let’s talk.