Adapt

Adapt is a company agent platform that runs AI agents on one shared runtime with connected tools, shared knowledge, model routing and permissions. It is open source, available locally, in the cloud or on-prem, and is currently a waitlisted developer preview.

Adapt is a platform for building and running AI agents at company scale. Instead of treating each agent as a separate project with its own infrastructure, it puts every agent on a single runtime that shares tools, knowledge and permission controls. The pitch is that adding another agent should not add another layer of operational complexity. The product is organised around five building blocks: one runtime, connected tools, a shared company knowledge layer, adaptive model routing and multiple work surfaces. Tools connect once through API, MCP or CLI, with identity and permissions inherited from the company's identity provider. Knowledge and past conversations accumulate into shared context and reusable workflows. Model routing chooses per task between frontier models, open-weight models and the customer's own keys, and inference can be brought in without markup or bought from Adapt. Teams reach their agents through Slack, the web, email and iMessage, with further channels added through plug-and-play surfaces. In practice, the site shows agents built from templates for jobs such as turning issues into pull requests, triaging support tickets against Shopify data, reconciling invoices on a schedule and sending a meeting brief before each call. Each layer is swappable: the model, the agent harness (Codex, OpenCode, pi, the AI SDK or Adapt's own), the sandbox (Modal, Vercel, Daytona, Browserbase or a private cloud) and the memory provider (Adapt, Mem0, Supermemory or Glean). Agents, skills, connections and permissions stay in place when a layer changes. Deployment is offered three ways: a free, open-source local option on the customer's own hardware, a hosted Adapt Cloud, and an on-premises enterprise version inside a customer VPC. At the time of the listing the product is a developer preview with a waitlist, so it suits engineering-led teams and enterprises evaluating a unified agent layer, rather than individuals looking for a ready-made assistant. It sits closer to agent infrastructure than to single-purpose automation tools.