New tools for self-hosting and managing AI agents

Managing coding agents, tracking their context across machines, and preventing outdated code are new challenges.

New tools let you run, sync, and correct agents locally or on your servers. Here is a look at what is available for self-hosting and managing AI agents.

Self-hosted Claude Managed Agents

Anthropic added a self-hosted option to the Claude API. You run the Managed Agents service on your infrastructure. You store the agents, sessions, and session history locally on your servers.

The local server sends model requests to a provider you choose. Anthropic is the default provider. Organization admins can turn off web fetch for their organization in the console settings. With web fetch turned off, requests that include the tool fail with a 400 error. The web fetch tool follows domain restrictions set by the organization.

The API update also introduced Claude Fable 5.2 and Claude Mythos 5.2. Both models have a 1M token context window and 128k max output tokens. They read other models' thinking blocks on the Claude API, allowing a conversation to switch models and keep its reasoning.

AgentRun: A managed platform for open-source agents

Agents often rediscover the same API and reread the same docs for repeated tasks. AgentRun is an open-source platform built by GREP.AI that adds durable execution to agent harnesses like Claude Code, Codex, OpenCode, or pi.

At the end of every run, the agent proposes notes, skills, or workflows to keep. Code checks the proposals, a grader judges them, and AgentRun commits the passing items to the agent's git folder. The next run starts with this learned context.

Every run lives in a single store. If you stop the server mid-run and start it again, the run carries on. You can give agents specific environments like a Linux sandbox, a desktop, an Android phone, or a browser.

You can configure tools to ask for approval first, and the run waits in the console until you answer. You configure your model keys in the project settings. You can run the server and one runner locally on port 7340 with this command:

npx @parcha/agentrun

Atuin Threads: Syncing agent sessions

Agent sessions store every command, output, and attempt you make while solving a problem. Atuin Threads syncs your agent sessions across machines.

You can start a task in Claude Code on a development server and finish it in Codex on your laptop. Atuin uses an end-to-end encrypted sync engine. It works with Claude Code, Codex, Pi, Omp, and OpenCode.

Agents can search and reference previous sessions using the Atuin Model Context Protocol. They can find past work and understand decisions.

You can install Atuin using this command:

curl --proto '=https' --tlsv1.2 -LsSf https://setup.atuin.sh | sh

You run atuin resume to access the interface. To give your agents access to the history, you install the protocol server:

atuin mcp install

Guide: Modernizing your coding agents

Coding agents frequently write bloated JavaScript for tasks that modern browsers handle natively. Modern Web Guidance injects modern platform expertise into coding agents.

The tool provides targeted guidelines that help agents use modern features like CSS highlights, dialog backdrops, and auto-sizing textareas instead of legacy libraries.

1. Install Modern Web Guidance using the interactive wizard in your terminal.

npx modern-web-guidance@latest install

1. Search for relevant guides directly from the command line without installing.

npx modern-web-guidance@latest search "animate a dialog modal backdrop"

1. Retrieve a specific guide by its identifier to feed into your agent.

npx modern-web-guidance@latest retrieve "animate-to-from-top-layer"

1. Scan your current codebase to find legacy patterns that the tool can help modernize.

npx modern-web-guidance@latest legacy-scan .

The catch

Self-hosted Claude Managed Agents are not available to everyone by default. You must contact your Anthropic account team to request access.

AgentRun requires Node 24.22.3 or newer, and it only runs on macOS or Linux. If Docker is running, new agents get a Docker sandbox, but the image takes a few minutes to build on the first use. If you use remote computers like Daytona or Browserbase in AgentRun, billing goes through your own account with those vendors.

The Atuin Model Context Protocol only supports a specific list of agent harnesses. If you use an unsupported harness, the agent cannot query your synced session history automatically.