Agents on Nullspace

Nullspace gives agents isolated compute, files, networking, terminals, and desktop environments. There are four integration paths — the right one depends on where the reasoning loop runs and whether Nullspace should operate a named project for you.

Local MCP

Your coding agent stays on your machine (Codex, Claude Code, Cursor, VS Code) and drives Nullspace through MCP tools.

Agent workspaces

The coding agent itself runs inside a sandbox — isolated filesystem, package installs, terminal, git, optional desktop.

Agent deployments

Your agent project becomes a named, versioned job or service that Nullspace runs, logs, and cleans up.

Framework integrations

Nullspace as a tool inside your app (OpenAI Agents SDK, LangChain) — or your framework app deployed onto Nullspace (Claude Agent SDK, LangGraph, CrewAI).

Which path fits?

You want to…	Use	Tradeoff to know
Let a local coding agent create sandboxes, run commands, and fetch files	Local MCP	The agent stays on your machine; only the work it launches runs sandboxed.
Put the coding agent itself in an isolated Linux workspace	Codex, Claude Code, Amp, or OpenCode	Provider credentials enter the sandbox environment.
Operate a repeatable agent app with runs, logs, outputs, and cleanup	Agent deployments	The app must fit the deployment contract: bundle, install, entrypoint, env, mode.
Expose sandbox operations as one tool in a larger agent app	OpenAI Agents SDK, LangChain, or bring your own	Your app owns retries, tool schemas, and framework state.
Run a framework app as the managed process itself	Claude Agent SDK, LangGraph, or CrewAI	Nullspace owns runs and sandboxes; the framework owns sessions and checkpoints.
Give an agent a browser or GUI	Desktop environments	—
Let a human inspect the same sandbox	PTY, SSH, or retained sandboxes	—

Two notes that save confusion later:

Live deployment modes are job and service. Other mode names you may see referenced (worker, triggered, workspace, interactive) are reserved for future models — they are not live.
Nullspace run IDs are not conversation sessions. Framework session, thread, checkpoint, and flow state stays app-owned. Use volumes, retained sandboxes, hibernate/auto-resume, or an external database when state must outlive one sandbox process.

The object model

Whichever path you pick, the same building blocks apply — see Agent primitives for the full model:

Primitive	What it means
Sandbox workspace	The isolated VM: commands, files, networking, PTY, SSH, desktop, volumes.
Agent project	A local directory with your app, `nullspace.agent.toml`, and an entrypoint.
Deployment	A named, uploaded project runnable as a `job` or `service`.
Run / service instance	One concrete execution, with logs, status, outputs, and a backing sandbox.
Framework state	Sessions, threads, checkpoints, memory — owned by your app, not Nullspace.

Start with an example

Minimal job

Deploy and run a small project end to end.

HTTP service

A long-running agent service with a public URL.

Workspace checks

Clone a repo in a sandbox and run checks like an agent.