v1.3.2
May 18, 2026- ·Recover agent sessions from prompt-too-long; route /compact natively
An IDE for the AI coding era — built for developers who pair with agents all day.
Every project · Every agent · One workshop
Hermes is a desktop IDE built for the way you actually work with AI: every project open at once, each session running its own agent. Claude is native via Agent mode — thinking, tool calls and diffs render as first-class UI, not terminal text. Codex, Gemini, your shell, your own scripts run alongside it in terminal mode. Git worktrees are first-class so a fix branch and a feature branch can coexist without a stash dance.
Each open session has its own model, mode, permissions, cwd, and history. No shared global state — switch sessions, switch context.
Agent mode talks to Claude directly via stream-json. Thinking blocks, tool calls, and diffs render as first-class UI — not terminal text.
Source-available under BSL 1.1. Converts to Apache 2.0 three years after each release. Free for everyone, forever.
Every other AI editor wants you to commit to one project at a time. Open a window, open a repo, ask the agent for help, close the window when you're done. Then do it again for the next project.
Hermes runs them all at once. The sidebar is a session list, not a file tree. Each session has its own model, permission mode, shell, working directory, and conversation. You tab between them like apps in a workspace — because that's what they are.
And worktrees are first-class. A fix branch and a feature branch can coexist as two sessions on the same repo, isolated on disk, sharing nothing. Switch between them at the speed of ⌘1, ⌘2 — no git stash dance, no "please commit your changes before checking out branch".
Each session has its own model (sonnet / haiku / opus), permission mode (manual / assisted / autonomous), and cwd. They don't share state.
Pin attached paths, share themes, browse a unified file tree, navigate every session with ⌘1..⌘9. You're not running nine apps — you're running one workshop.
Spawn a session on a new worktree — isolated branch, isolated cwd, isolated state. Merge it back when you're ready. No stash dance.
Agent mode is a per-session subprocess that streams Claude's output as structured events — not bytes on a TTY. Thinking, tool calls, file diffs, permissions: every part gets its own first-class UI.
You can interrupt mid-thought. You can change the model mid-turn. You can hand a permission request the file it needs without cancelling and restarting. The terminal couldn't do any of that.
Stack roles. Pick a style. Drop your task in. Hermes compiles a clean, structured Markdown prompt before sending it — so every session starts from the same high water mark instead of a blank line.
A Backend Engineer who's also a Security Auditor is one click apart from a Backend Engineer alone. Roles add together; Hermes resolves the seams.
Concise isn't a switch — it's a level from 1 to 5. Pick the loudness of every voice the agent should answer in.
Once you've composed a good prompt structure, save it as a template. Future sessions pre-fill the same shape — change just the task.
Every theme keeps the two-voice contract — a warm operator and a cool agent — except the three mono-tone themes that alias both voices to one accent. Choose a temperament, not a palette.
Every tool call asks first. Every diff lands in front of you. Slow, exact.
Read-only tools auto-approve. Writes still ask. The default rhythm.
Bypass mode — let it run. Mark its work with the red Permission chip so you remember.
Hermes is designed around not touching the mouse. Every common action has a one-handed shortcut.
Free. Source-available. macOS, Linux, Windows. Auto-updates as we ship new editions.