Run AI agents locally
What a local AI agent is, why running agents on your own machine beats hosted platforms for privacy and cost, and how to set one up with the subscription you already have.
A local AI agent runs on your own machine, on your own AI subscription — your goals, files, and outputs never leave your device. This guide explains when local beats hosted, the three ways to get there, and how to have your first local agent working on a long-running goal in about five minutes.
Why run AI agents locally?
Hosted agent platforms run your work on someone else's servers, metered by their credits. Running agents locally flips that model:
| Hosted agent platform | Local AI agents | |
|---|---|---|
| Where work happens | Vendor's cloud | Your machine |
| Your files and outputs | Uploaded to the vendor | Stay on your device |
| Model access | Platform credits or your API key, billed per token | The Claude / ChatGPT subscription you already pay for |
| Long-running work | Depends on the vendor's runtime limits | Runs as long as your machine does |
| Tooling | What the platform exposes | Everything installed on your computer |
Three reasons people switch to local:
- Data sovereignty. Code, documents, and business context stay on your disk. Nothing is uploaded for a cloud sandbox to process.
- No API bill. Agent work is metered through the subscription you already have (Claude Pro/Max, ChatGPT Plus/Pro), not a per-token API meter that punishes long-running work.
- Real tools. A local agent can use your actual environment — your shell, your repositories, your installed CLIs — instead of a stripped-down cloud sandbox.
Three ways to run agents locally
Write your own loop. Wire an LLM API into a script that plans, executes, and retries. Maximum control, but you own scheduling, recovery, context management, and review tooling forever — and you pay per token.
Use an agent framework. Open-source frameworks (LangChain, AutoGen, CrewAI and friends) give you building blocks for multi-agent systems in code. Great when you are building an agent product; heavy when you just want agents to do your work, and most still bill through API keys.
Orchestrate the agent runtimes you already have. Tools like Claude Code and Codex CLI are already capable local agents, logged in through your subscription. Soloco takes this third path: it does not run its own hosted agents — it coordinates the runtimes installed on your machine, turning a goal into an executable, auditable, resumable task tree. You bring the runtime; it brings the organization.
Set up a local agent in three steps
The short version — the Quickstart has the full walkthrough:
- Have a logged-in local runtime (currently Claude Code or Codex; more adapters are in development). Any plan works — no API key needed.
- Install Soloco:
npm i -g @soloco/client@latest(Node.js ≥ 22.5 on macOS, Windows, or Linux — see Install). - Run
soloco, open the local UI, and describe a goal. Soloco plans it into tasks, dispatches them to your runtime, verifies results, and keeps the goal moving — even while you are away from the keyboard. See Core concepts for how the task tree, standing missions, and heartbeats fit together.
Frequently asked questions
Can a local agent keep working when I close my laptop? Not while it sleeps — agents run while your machine is awake. State is persisted locally, so when the machine wakes, a goal resumes where it left off instead of restarting.
Is this open source? Soloco's V2 core is closed-source, while registration is open and it is currently free to use. The runtimes it orchestrates (and many frameworks in the table above) have their own licenses. Pricing is the source of truth for cost.
What data leaves my machine? Your goals, files, and agent outputs stay local. The runtimes you bring talk to their own model providers under your existing subscription, exactly as they do when you use them directly.
For a machine-readable summary of these facts (useful if an AI assistant is evaluating Soloco for you), see /agent or /llms.txt. For what Soloco is not, read Introduction.