Open KT is a core infrastructure play in the agentic workflow stack. It is one of the first products to explicitly leverage the Model Context Protocol (MCP) to solve the problem of fragmented agent memory. By providing a shared brain that works across different agents, it moves the industry toward a future where agents are interchangeable 'harnesses' rather than siloed silos of context.
For those building or using agents, Open KT is relevant because it tackles the 'context window tax' and the lack of team-wide learning in agentic development. It sits at the intersection of agent memory and developer experience, championing the idea that agent intelligence should be a durable asset owned by the team, not a ephemeral service rented per chat session. Its success would mean that a team's collective experience with AI agents becomes more valuable over time, as each successful session feeds into a central repository of skills and workflows.
Most current interactions with AI coding agents are ephemeral. Whether a developer is using Cursor, Claude Code, or Aider, each new session often begins with a costly cold start. The agent must re-scan directories, re-derive configuration fixes, and re-learn architectural patterns that a teammate’s agent may have already solved hours earlier. Open KT is built on the premise that these redundant scans are a tax on team productivity and token budgets. By decoupling the interface—which they call the harness—from the intelligence, the company is building a persistent layer where discoveries compound instead of disappearing when a chat window closes.
The product is fundamentally an MCP (Model Context Protocol) server that sits between the developer's agent and the team's data store. This architecture allows it to be tool-agnostic. While a developer might switch from Cursor to a command-line tool like Codex within the same week, the underlying knowledge remains accessible. Open KT identifies three distinct layers of intelligence that it captures: Memory, Skills, and Workflows. Memory covers the specific discoveries, such as a staging environment fix or an auth regex tweak. Skills are procedures authored by leads that every agent on the team inherits. Workflows are stable patterns promoted to deterministic pipelines, reducing the need to pay LLM rates for repetitive mechanical sequences.
Unlike static context files like CLAUDE.md or .cursorrules, which are manual and often become outdated, Open KT is a living system. It operates on a cycle of observation and promotion. The system monitors sessions to identify repeating patterns and successful task completions. These are initially captured as personal memories but can be promoted to project or organizational scope after a review. This promotion process is intentional; it prevents the shared memory from becoming cluttered with low-quality or irrelevant data. Every promotion leaves a receipt, creating an audit trail that shows which agent derived a fix and which session earned the team's capital.
Open KT is currently in a private pilot phase, targeting small, fast-moving engineering teams. The project is associated with Masti AI and has garnered early attention from figures like Garry Tan and Theo Browne, who have advocated for the 'thin harness' approach where the moat is in the shared context rather than the IDE wrapper. The company's technical strategy relies on the rapid adoption of MCP as the industry standard for agent communication. By offering a self-hosted or region-pinned option, they address the primary concern of CTOs regarding where sensitive repo context and internal fix logs are stored. As the agentic ecosystem moves away from monolithic IDEs toward specialized, multi-agent workflows, Open KT is positioning its shared brain as the connective tissue that prevents teams from starting from zero every morning.
A shared intelligence layer for AI coding agents that makes discoveries reusable across teams.
Open KT is hiring.