Forkbench for Claude Code
The Mac terminal built for Claude Code
Claude Code is Anthropic's coding agent that runs in your terminal, reads your repository, edits files and runs commands.
Forkbench is a Mac app for running Claude Code where the credentials it needs are used by the command and never handed to the agent, and a Thread can be locked to its folders by the macOS kernel so nothing it starts can read outside them. Around that, each session gets a real login shell in its own tab and a Thread holding one job's folder, notes and keys.
What breaks in a plain terminal
One Claude Code session in a plain terminal is a good experience. Four of them is where it frays. The tabs all look alike, so you lose track of which repo each one is in. Two sessions edit the same file and undo each other. The plan Claude spent twenty minutes building lives in scrollback you will never scroll back to, and when the context window rolls over or you close the pane, it is gone. And the moment Claude needs to push, deploy or call a paid API, the honest options are pasting a token into the prompt or exporting it into a shell the agent can read.
How Forkbench runs Claude Code
Keys it can use and never read
Store the token in Vault once. When Claude needs to push to GitHub or call an API, Claude runs that one command and takes the result. The key never appears in the prompt, the command line or the transcript. Optional Touch ID on every use.
Folders it cannot wander out of
Lock a Thread to the folders the job is about. Every shell in it starts under a macOS Seatbelt profile that the kernel enforces and every process inherits, including build scripts and a dependency's install hook, so a read or a delete outside them is refused rather than discouraged. Builds still run, paths still resolve, and the network is untouched.
A real login shell, not a wrapper
Every tab is a genuine PTY, so claude launches exactly as it does in any terminal, with your prompt, your aliases, your shell config and your MCP servers untouched. Forkbench does not proxy the agent or sit between it and Anthropic.
Sessions that stop looking alike
Group related tabs into a Thread. The Thread carries its folder, the notes you granted it and its own secrets, so switching from the auth work to the billing work is switching Thread, not hunting for the right tab.
A plan that outlives the context window
Claude writes tasks, progress notes and blockers to the Thread's board rather than into scrollback. The board is written to disk as it changes, so a compaction, a closed pane, a quit or a reboot leaves it intact. A new session reads it and picks up where the last one stopped.
Its own branch, in one click
Give a session a fresh worktree from the git panel so it cannot touch your working branch. The panel always reflects the tab you are focused on, and Forkbench never merges, rebases or changes your git on its own.
Hand its work to Codex tomorrow
Nothing on a task records which agent wrote it. A task Claude Code planned this morning is claimable by a Codex or Gemini CLI pane this afternoon, with the notes and the discussion that settled a question still attached.
Answer it from your phone
Claude blocks on a question while you are away from the desk. Talk attaches to the conversation already open on your Mac from any browser, so you answer and it keeps going. It attaches to a live session, it cannot start one. Part of Pro.
Getting Claude Code running
- 1Install Forkbench, open a Thread on your repo, and run claude in a tab exactly as you would anywhere else.
- 2Add a second tab for a second session, and give it a fresh worktree from the git panel so the two never touch the same files.
- 3Give the Thread an objective. Claude adds tasks to the board as it plans, and claims one before it starts work on it.
- 4Grant the Thread the notes it should read and the secrets it may use. Everything is deny by default until you grant it.
- 5Close the pane whenever you want. The claim releases immediately and the plan stays on the board for the next session.
What this does not do: Forkbench does not make Claude Code better at coding, and it does not replace your Anthropic subscription. It runs the same binary you run today. What it changes is everything around it: where the work is recorded, what the agent can reach, and what is still there tomorrow.
Claude Code questions
Does Claude Code run normally inside Forkbench?
Yes. Every Forkbench tab is a real login shell, so claude starts the same way it does in Terminal, iTerm2 or Ghostty, with your shell config, aliases, hooks and MCP servers exactly as you set them up. Forkbench does not wrap the agent or route anything through its own servers.
Can I run several Claude Code sessions at once?
Yes. Each session gets its own tab, and you can give any tab a fresh git worktree in one click so two sessions never edit the same working tree. Related tabs group into a Thread that scopes their folder, notes and secrets, and the Thread's board leases each task to one agent so two sessions never pick up the same work.
What happens to Claude's plan when the context window compacts?
Nothing, if the plan is on the board. Tasks, progress notes and blockers are written to a per-Thread file on your Mac as they change rather than living in the conversation, so compaction, a closed pane, a quit or a reboot leaves them intact. A fresh session reads the board and continues.
How does Claude Code use my API keys without seeing them?
Through Vault. The secret stays on your Mac, and when the agent runs a command that needs it, the command gets what it needs and the agent does not. The agent receives the command's result and never the raw value, so it stays out of prompts, logs and the transcript. It is a seatbelt, not a guarantee: an agent you deliberately authorise for a command could still print what it was handed. And Forkbench does not stop Claude Code from reading a .env file already sitting in the project, because it has the same filesystem access you do. Move the key into Vault and there is nothing left in that file worth reading.
Is this a replacement for Claude Code's Agent Teams?
No, and they are not competing for the same job. Agent Teams runs Claude sub-agents inside one Claude session. Forkbench is the app that session runs in, so Claude Code sits beside Codex, Gemini CLI and your own scripts on one board. Use Agent Teams inside a Forkbench tab if you like.
What does it cost to run Claude Code in Forkbench?
Nothing extra for the agent itself. You keep running Claude Code on the Anthropic subscription you already pay for, and Forkbench never bills per request or meters AI credits. Forkbench is free to download, and the terminal, your notebooks and a board on every Thread stay free. Vault and running more than one agent on a board are Pro, and every account starts with 30 days of it without a card.
Running something else?
Forkbench is agent-agnostic, so any CLI agent runs in a tab. These have a page of their own: