A harness of its own.
Condensate runs the model-and-tool loop, manages context, and records the work. Choose a supported provider model for the task.
Native harnessAn agent runtime, built to connect
Your agents.
Working together.
Give a task to an agent. Bring in help. Change direction. Come back to the work.
Condensate keeps it connected.
Developer preview. Running locally requires repository access.
See the idea in motion
One task. A shared direction.
A guided product walkthrough.
Illustrative data; no live agents or model calls.
WorkspaceInteractive example
Start with a task.
The request becomes part of a durable thread. The agent can inspect the project and begin its work.
message acceptedRecorded in the threadturn runningModel and tools beginFollow the work, or send a new direction…
This walkthrough explains the runtime’s task, child-thread, messaging, and history capabilities. Application interfaces and integration support vary.
From an answer to ongoing work
An agent needs more
than a model.
A model can reason about a task. A harness gives it context and tools. A runtime keeps the work organized as tasks grow, agents collaborate, and you change direction.
Condensate brings these pieces together. Use its own harness, run Claude through its SDK adapter, or connect an existing harness through tools and hooks.
Explore the three integration pathsRoom for other harnesses.
The Claude adapter integrates SDK sessions. Middleware supplies MCP tools and policy hooks to compatible external harnesses.
Tools and hooksYour product on top.
Use the HTTP API and TypeScript SDK to build a chat interface, a review workflow, or another application around agent tasks.
Build with the SDKKeep the thread
The work goes on.
The context can change.
A long task can outgrow a model’s context window. Condensate records its history and creates a smaller view for the next model call.
You can inspect earlier events even when the model needs to reread a detail. The history and the next prompt serve different purposes.
See how context windows workDURABLE HISTORY Retained
RequestReadsToolsDecisionsRecent
Context projection
NEXT MODEL INPUT
SummaryRecent workRoom to continue
Different surfaces. Shared foundations.
See what’s built with it.
Dex
↗Find, follow, and continue your agent tasks in a browser.
Task interfacegt-review
↗Bring repository changes and review conversations into one place.
Code reviewDojo
↗Build learning around a tutor and a workspace suited to the subject.
Learning surfaceThese applications reuse Condensate packages in different ways. Their guides explain setup and session ownership.
Start with one task
Make it work
for you.
With repository access, run the runtime locally, connect your account, and open Dex. The public docs explain the APIs and architecture for your own application.
Follow the setup guideFor developers with repository access
git clone https://github.com/condensate-dev/condensate.git
cd condensate
bun install --frozen-lockfile
bun run --cwd packages/agents-host startThe repository is currently private. Requires access and Bun; account and workspace setup follow in the guide.
Read the SDK examplesA few things to know.
How can I try it?
Explore the interactive walkthrough and public documentation here. Condensate is a developer preview; running it locally currently requires access to the private repository and a supported model account.
Is Condensate a model?
No. It supplies the harness and runtime around models: tools, task history, execution, and coordination. You connect the provider accounts and models you want to use.
Can I keep using another harness?
Compatible external harnesses can connect to middleware through MCP and supported lifecycle hooks. They keep their own agent loop. Available controls depend on the client integration; MCP alone does not transfer ownership of a session.
Does the demo run real agents?
The walkthrough uses illustrative data and runs entirely in your browser. Follow the setup guide to run Condensate with your own model account and workspace.
Can I build something beyond coding?
The runtime exposes general task and event APIs. Dojo is one example of an application that uses shared engine packages for learning. Your application supplies the interface, tools, and workflow for its domain.