All postsSHORT2 min read
Self-Hosted Claude Agents Can Overwrite Your Branches
Self-hosted Claude Managed Agents restricts push branches but doesn't refuse force pushes. Protect branches on your Git host and configure tracing.
Self-hosted Claude agents can overwrite any branch you let them push to. Anthropic's docs say its self-hosted server doesn't refuse force pushes. I'd check branch protection on the Git host before giving an agent access.
What Anthropic announced
Anthropic announced self-hosted Claude Managed Agents on Nov 12, 2026. It's in private beta. Contact your Anthropic account team to request access. The announcement and docs don't list a price.
The whole service runs in your own cloud or on-prem, including the harness, sandboxes, credentials and session history. Model requests go where you choose: Anthropic directly, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry or your own gateway.
Listed branches still need protection
The self-hosted differences page says agents can push only to branches you list. That list is a guardrail, but the server doesn't refuse force pushes. An agent can overwrite any branch you've listed. To refuse force pushes, protect the branch on your Git host.
The Git token your client sends never reaches the sandbox. Git traffic goes through your server, which adds the token. This Git-specific behavior is separate from Anthropic's broader statement that vaults are designed to keep credentials out of the sandbox.
There's also a verification catch: the server's log can say ok even when your Git host refused a push. Check the branch on the Git host to see whether the push succeeded.
You run and secure the service
The self-hosted docs say you operate the server, its database and the workers where agents run shell and file tools. You run and secure the server and sandboxes, and install each release yourself.
The agent's bash tool is a real shell with the worker process's permissions and network access. The worker doesn't isolate that shell. Isolation comes from the container, VM or other sandbox that the worker runs in.
For a self-hosted sandbox, set traffic limits in your own firewall. My recommendation: review the sandbox's firewall rules alongside the branch permissions before handing the setup to a team.
Session history and tracing
Session history is stored in your own database. The self-hosted docs list PostgreSQL 16 or later. Sessions and their history stay in your database until you delete them.
Traces and metrics are off until you turn them on with the standard OpenTelemetry environment variables. Configure those variables to send telemetry to your monitoring. Keeping session history in the database and enabling tracing are separate parts of the setup.
What goes to the model provider
Each model call sends the conversation so far, including tool inputs and results, to your inference provider. That data flow still needs a place in the team's review when the agent service runs on its own infrastructure.
The builder angle (my opinion, no income guaranteed)
I'd sell teams an agent lockdown check. The deliverable: branch protection on the Git host for the branches an agent may push to, a short firewall list for the sandbox, and tracing turned on and sent to their monitoring.
I'd check the branch on the Git host when verifying a push, then confirm the configured tracing arrives at the team's monitoring. That's a proposed service, with no income guaranteed.
Daily email
Get each breakdown in your inbox.
One email each evening, right after the Short goes live: the AI tool it covers, the full write-up with sources, and how to get paid for it. Free, and you can unsubscribe anytime.