Tulina vs. Dust
The context layer that writes itself.
Dust gives teams a workspace to build AI agents. People with builder roles configure each agent, point it at synced documents, and revise it as the company changes. The agents are the product, and your team maintains them.
Tulina starts from a different belief. The understanding your AI works from should build itself, stay current on its own, and follow your team into whatever AI they already use. Nobody should have to become a builder first.
Dust
Tulina
Operating model
People build and maintain a fleet of agents
One company brain that builds itself
Company context
Retrieved from synced documents at query time
Distilled into a living map, sourced and dated
Retrieval
Semantic search returns matching chunks
Routing in the company brain sends each question to the pages it needs
Context upkeep
Docs sync; the instructions on top are revised by hand
The brain revises its own pages
Self-improvement
Nightly suggestions await a person's approval
Automatic, every change sourced and dated
Memory across runs
Agents start fresh each time
Every run folds back into the context
Surfaces
Its own surfaces; an MCP server that stops at its agents
Inside Claude, ChatGPT, or Mistral, your whole stack behind it
Company-wide rollout
Builders, champions, a staged rollout
Still one connection
Shared context.
Dust’s agents point at synced documents, and what each one understands lives in instructions a person writes and revises. Tulina’s company brain is the understanding itself: written by the system, sourced, dated, current, and shared by every agent that reads from it.
Shared connectors.
Dust can surface its agents over MCP, but the actions stop at Dust’s own tools. Tulina’s one connection carries your whole stack, so the same connectors work inside Claude, ChatGPT, or Mistral without being rebuilt for each.
Shared processes and databases.
When Dust learns from usage, it drafts a suggestion for a builder to review, one agent at a time. Tulina folds every run, correction, and decision back into the same shared context automatically, so what one process learns, every process already knows.