Introducing Deft: a workspace where humans and AI work together

[Maneek21]

Introducing Deft

Untitled.design.7.1.mp4

We built a new kind of workspace for teams working with AI. It is called Deft.

Deft brings conversations, tasks, knowledge, people and AI agents into the same workspace, so humans and agents can work from the same underlying context instead of constantly passing information between disconnected tools.

The idea behind Deft rests on two bets.

1. Humans and AI should work from the same record

Most AI agents today sit outside the place where work actually happens. The conversation might be in Slack, the task in Linear, and the knowledge in Notion. Then someone opens ChatGPT, Claude or Codex and reconstructs enough of that context for the agent to help.

We think that is backwards.

Humans and AI should work from the same record. An agent should be able to understand the conversation that created a task, the people involved, the relevant knowledge and the decisions that came before it. When an agent takes action, that work should become part of the same record rather than disappearing into another AI conversation.

In Deft, that can look like:

Conversation → context → task → agent → approval → action → receipt

Agents can read workspace context, participate in work and take governed actions. Sensitive actions can require approval, and completed actions leave receipts showing what happened.

The goal is not to add an AI sidebar to another productivity product. The goal is to make humans and agents participants in the same system of work.

2. The workspace should expand to fit the company

The second idea is about software itself.

Companies rarely stay within the boundaries of a handful of standard tools. Eventually they need a CRM, a recruiting workflow, an approvals system, a partner tracker, or some internal application that only makes sense for that particular company.

Today, each of those needs usually becomes another application with its own database, permissions model, search layer, integrations and copy of company context.

We think there is another possibility.

The workspace itself should be able to expand as the company changes. New applications should be able to inherit the environment that already exists around them, including people and identity, permissions, knowledge, search, agents, context and governance, rather than rebuilding those primitives every time.

In Deft, Modules are the first step in that direction. Contacts, for example, is implemented as a module that can use capabilities already provided by the workspace.

The goal is not to turn Deft into every piece of software a company could ever need. It is to create a foundation on which company-specific software can live without becoming another disconnected island.

Put together, those are the two core ideas behind Deft:

Humans and AI should work from the same record, and the workspace should be able to expand to fit the company around it.

What Deft looks like today

Deft is already a functioning alpha. It includes real-time chat, spaces, DMs and threads; tasks with Board, Table, Timeline, Calendar and Pipeline views; notes, company knowledge, decisions and references; calendar context; people, teams, roles, inbox and notifications; Defty, a built-in AI agent; personal MCP connections for Codex, Claude, ChatGPT and other MCP clients; agent employees; explicit scopes and trust levels; approval flows; action receipts; provider-neutral AI configuration; Modules, with Contacts as an early example; and Docker + PostgreSQL self-hosting.

Deft also works without an AI provider configured. That is deliberate.

AI is not the substrate. The shared workspace is.

Why we are putting it out now

Deft is still an alpha. There will be bugs, parts of the product will change, and some of the architecture will probably change too.

More importantly, both theses above now need to survive contact with people who did not build the thing.

If you are building with agents, MCP, open-source productivity software or internal tools, we would love for you to take a look. Clone it, self-host it, connect an agent, inspect how permissions and approvals work, build a module, or use it for something we did not anticipate.

Then tell us where the model breaks.

In particular, we want to learn whether working from a shared record actually makes humans and agents better collaborators, and whether an extensible workspace can meaningfully reduce the need for another disconnected application every time a company develops a new requirement.

That is Deft.

And this is the beginning of the experiment.

GitHub: https://github.com/Maneek21/Deft
Website: https://deft.ing/