input.received01
model.requested02
tool.completed03
turn.completed04
Measure from step to task.
Trace model steps, agent turns, and completed tasks—from input and tool calls through to final output—so every improvement is grounded in evidence.
1import { Agent } from "@nylorun/harness"2import { Runtime, serveAgents } from "@nylorun/runtime"3import { Hono } from "hono"4import { gcp, github, knowledgeGraph, sandbox } from "./capabilities"5 6const wiki = Agent({ id: "wiki", name: "Generate Wiki for Repo" })7 .use(github)8 .use(sandbox)9 .use(knowledgeGraph)10 .use(gcp)11 .build()12 13const app = new Hono()14app.route("/agents", serveAgents({15 agents: [wiki], runtime: new Runtime(), basePath: "/agents",16}))Build agents you can measure, optimize, and evolve.
Harness keeps execution visible and composition explicit, while you remain in control.
Harness
provider
observer
transport
Optimize on your terms.
Serve agents through Hono wherever Hono runs. Bring the sandboxing, durability, and observability you need, while retaining complete control of your data, credentials, and deployment.
Agent
instructions
capability.use()
model
Evolve through composition.
Package instructions, tools, model selection, and context into reusable capabilities. Compose your own or third-party middleware with .use() to extend an agent without rebuilding it.
One simple loop. Powerful agent execution.
Harness brings models, capabilities, and tools into one visible loop that continues until the agent reaches an answer.
InputModelToolsOutput
task input
Model
Request. Response
Tools
Execute. Results
In-ProcessSandboxMCP
final output
task input
Model
Request. Response
InstructionsToolsHistoryContext
Tools
Execute. Results
In-ProcessSandboxMCP
final output
Build and run agents. One seamless stack.
Create, compose, run, and inspect agents with an open-source stack that works together—and keeps you in control.
Creatornew agent
Agent running
Agents
Create and run an agent from the starter in minutes.
@nylorun/create-agentgithub
sandbox
gcp
Agent.use()
Harness
Compose reusable capabilities into the agent you need.
@nylorun/harnessLocalsame agentCloud
Runtime
Run locally, then take the same agent to cloud infrastructure.
@nylorun/runtimeBuild a wikiWorking…
model
tool
done
Studio
Chat with agents and inspect their session logs side by side.
@nylorun/studioNylorun Harness, in plain terms.
The core decisions, package boundaries, and current availability in one place.
What is Nylorun Harness?
Nylorun Harness is an open-source, provider-neutral TypeScript library for building agents around a visible model-and-tools loop. It is designed for developers who want to compose agent behavior while retaining ownership of their application and infrastructure.
Read the Harness overviewDoes Nylorun Harness work with Hono?
Yes. Nylorun Runtime mounts agent routes in a Hono application, while Harness runs the in-process model-and-tools loop. Your application remains responsible for transport, authentication, provider accounts, and deployment.
Learn about deployment responsibilitiesWhat do Creator, Harness, Runtime, and Studio do?
Creator scaffolds a compatible project, Harness composes capabilities and runs the agent loop, Runtime provides the local host and CLI, and Studio is an optional local dashboard for inspecting agents and sessions.
See the package overviewWho owns the agent's data, credentials, and deployment?
You do. Harness and Runtime provide portable contracts and local tooling, while your application owns provider accounts, credentials, storage, transport, access controls, and deployment decisions.
Review production responsibilitiesIs Nylorun Harness available today?
Harness, Runtime, Studio, and Creator are available as experimental beta packages. The Managed Agents platform is in limited early access and is being developed separately from the open-source stack.
Start with the quickstartManaged Agents
We're building the managed agents platform and are working with limited set of partners.