Skip to main content
AgentFS lets you run AI coding agents like Claude Code, Codex, and OpenCode in a sandboxed environment where file changes are isolated from your original source tree.

Supported Agents

AgentCommand
Claude Codeagentfs run claude
Codexagentfs run codex
OpenCodeagentfs run opencode
Any CLI toolagentfs run <command>

Workflow

Here’s a typical agentic coding workflow. This example uses Claude Code, but the workflow is the same with OpenCode, Codex, and others.

Step 1: Clone the Repository

Clone the git repository to a host directory. Keep this source tree clean from build artifacts. When there’s a merge to main, pull to this repository to keep it up-to-date.
git clone https://github.com/tursodatabase/agentfs.git
cd agentfs

Step 2: Start an AgentFS Session

Use agentfs run with the --session flag to start a session. Pick a name that can eventually become your git branch name.
agentfs run --session fix-bug-123 claude
The agent now runs in a copy-on-write environment. Your original source tree remains untouched while the agent works.

Step 3 (Optional): Manual Investigation

If the agent does something unusual, open another terminal and join the same session to investigate:
agentfs run --session fix-bug-123 bash
The filesystem is shared between the shell and the agent. You can browse files, run tests, or check what’s happening while the agent works.

Step 4: Push to Branch

Once the agent has completed the work, ask it to create a branch, commit the code, and push:
“Create a branch called fix-bug-123, commit these changes, and push to origin”
You can then create a PR manually from the pushed branch.

Why This Workflow?

  • Clean source tree - Your original clone stays pristine and easy to update from upstream
  • Isolated changes - Each task gets its own session, preventing cross-contamination
  • Easy investigation - Join any session to see what’s happening in real-time
  • Standard git flow - Push branches and create PRs as usual

Next Steps