deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

JoeBro puts AI agents beside your files in a free open-source Mac workspace

JoeBro is a free GPLv3 Mac workspace that opens documents next to AI agents like Claude Code and Codex, with every file write gated behind a diff or an explicit approval.

JoeBro puts AI agents beside your files in a free open-source Mac workspace

A workspace, not a chat window

The developer behind JoeBro, announcing the project on dev.to, says it was built to invert the usual AI workflow of copying text out of a chat window and pasting it into the place where you actually work. JoeBro is a free, open-source macOS application, licensed under GPLv3, that keeps documents and code open and editable alongside the conversation. Agents — Claude Code, Codex and OpenCode, plus any model you point it at — read and change the same files you are editing yourself.

The built-in editor handles Markdown with a live outline for jumping around long documents, and also opens Word .doc and .docx files and spreadsheets. You bind a folder to a workspace, and the agent's file tools for listing, reading, creating and editing operate inside that folder.

According to the announcement, the app is local-first: chats and memory live in a single SQLite file, with no account and no telemetry. Installation is a one-line shell command, and the app requires macOS 26 or newer. The source code is published on GitHub.

How agent edits are gated

The developer identifies file editing as the riskiest part of giving an agent access to your work, and describes three layers of control, ordered by how often each matters.

First, if a document is open in the editor, an AI edit is queued as a pending diff that you apply or reject in place — no modal dialog, no timeout. The change shows up exactly where you are already looking rather than in an interrupting prompt.

Second, if the file is not open, the write is gated before anything touches disk. Where an approval channel exists, the edit waits for an explicit yes; where there is no live channel to ask on, the write is refused outright rather than applied silently.

Third, commands sit behind a separate switch. One setting raises a blocking approval before shell and Python tools run, on the reasoning that executing code is a different class of risk from editing a document.

Engineering around model failure modes

Long files get two specific fixes in the file tools. A long document can be created in chunks, with the server appending each incoming part to what is already saved and re-inserting the blank line that gets trimmed from tool arguments — a single newline for CSV-style files — so two chunks do not run together.

Truncation is detected rather than misreported. When a reply hits its token limit partway through an edit, the tool arguments stop being valid JSON; parsing that as an empty argument set makes the tool fail with a misleading missing-path error. The tool layer instead checks for non-empty, unparseable arguments and treats them as a truncated call.

A small alias table also maps the parameter names models invent — file, filepath and file_name all resolve to path, and old_string to find — because discarding those calls would waste an entire turn.

Terminal agents, shared memory and an iPhone app

Claude Code, Codex and OpenCode run in their own terminals. JoeBro displays their sessions in the same window as your files, so you can read what they are doing, respond to their permission prompts and see the files they change — from the Mac or from an iPhone app. The agents also share a single, per-project memory over MCP.

The developer is openly soliciting feedback on one design choice: whether queuing edits as diffs should also be the default for files that are not yet open. Today those writes go through the approval prompt where one exists, and the author writes of going back and forth on whether to open such files in the editor automatically instead.

Why it matters

Most AI assistance still happens in a separate chat window, with the human acting as courier between the model and the workspace. JoeBro is a concrete attempt to make the file itself the shared surface between person and agent — and its permission model is the interesting part: visible diffs for open files, refusal rather than silent writes for closed ones, and a distinct gate for command execution. That layered design is a reusable pattern for anyone building agents that touch a real filesystem.

The unglamorous details matter too. Handling truncated tool calls and invented parameter names is exactly the plumbing that decides whether agentic tools feel reliable day to day. With a GPLv3 license, SQLite-local storage and no telemetry, JoeBro also positions itself as a privacy-conscious alternative to hosted AI workspaces — though every claim here comes from the developer's own announcement rather than independent testing.

  • #open-source
  • #macos
  • #ai-agents
  • #developer-tools
  • #local-first

Related posts