Desktop · macOS & Linux · v0.3.1

codrift

Run many AI coding agents at once without losing track of one.

Claude Code, Codex, Opencode, Gemini, Copilot, Cursor and raw shells, against every repo in a project at the same time.

Live terminals, a branch per directory, one diff across all of it, and a memory store the agents share.

Source on GitHub

Four agents, three repos, one window. Live, from a real session.

j k hold a lane still, the same keys that move the sidebar cursor in the app.

01 / floor

Every agent, in its own terminal, on screen at once.

Codrift does not wrap your agents or paraphrase them. Each one runs in a real PTY and renders byte-for-byte in an embedded terminal, so Claude Code looks like Claude Code and Codex looks like Codex. Split the pane and watch two of them at the same time.

Two Claude Code agents running side by side in Codrift: the left pane is proposing a Go edit, the right has just finished writing a failure-modes document — ⌘D splits the content area. Each half keeps its own initiative, agent and view.

Seven adapters, one launcher

Claude Code, Codex, Opencode, Gemini, Copilot, Cursor, and a raw $SHELL. Start any of them against any directory; run several in the same one if you want a second opinion.

Sessions outlive the window

Claude sessions are persisted and resumed. Quit Codrift, come back, and the agent picks up the conversation instead of starting over.

02 / attention

Output stopped being scarce. Your attention didn’t.

Four agents will produce more work than you can read. The job is no longer typing, it is routing yourself to the one that is actually stuck. A row of terminal tabs cannot tell you which: a tab title carries the repo name, never the fact that an agent stopped ten minutes ago to ask something. So you check all four, every time. Codrift makes it a glance: every agent carries a state, the sidebar rolls them up per initiative, and anything waiting on a human says so in words, not just colour.

running

Working. Leave it alone.
needs input

Asked you something. This is the only one that costs you time.
idle

Finished its turn, holding the session.
stopped

Exited or crashed, with its exit code.

Codrift sidebar showing an initiative with 3 waiting, while an agent asks how the usage endpoint should carry two rate-limit scopes — “3 waiting” on the initiative row. You go straight there.

03 / conductor

Hand the plan to an agent and let it run the others.

Codrift ships an MCP server. Point one agent at it, give it the initiative, and it plans the work, starts sub-agents in the other directories, sends them instructions, reads their output, and writes what it learns to the shared memory. There is no glue code and no orchestration DSL: the conductor is just an agent with tools.

Press o to start orchestration on the selected initiative, or wire the MCP server into your own tooling with codrift mcp install.

orchestration.md, read by the conductor

# Rules for workers

- Read the initiative memory
  before touching anything.
- Write a decision entry the
  moment you pick between
  two approaches.
- Never widen the public API
  without saying so in memory
  first.
- Stop and report if a change
  would need a migration.

## Merge order

edge -> api -> dashboard

04 / isolation

Six agents. Six branches. Your checkout untouched.

Turn on worktrees for a directory and Codrift checks it out on codrift/{id}/{slug} before any agent starts. Parallel agents stop colliding, and nothing lands in your working tree until you decide it should. The diff view shows every change across the whole initiative in one scroll, so review is one pass, not six.

Codrift diff view showing syntax-highlighted additions and deletions in lib/atlas/rate_limit.ex with a plus thirty, minus two summary — 2 shows every changed file across every directory, highlighted.

05 / memory

The context you keep re-typing, written down once.

Every initiative carries a searchable SQLite FTS5 store. Agents read it before they start and write to it when they decide something. The next agent inherits the reasoning instead of rediscovering it, which is cheaper in tokens and much cheaper in contradictions.

decision claude/atlas-api

Bucket key is {api_token, route_class}, not api_token alone. A read loop must never starve a payout. Rejected: one global bucket per token (simpler, but it couples the two failure modes).

snippet claude/atlas-api

defp encode({token, class}), do: "rl:#{class}:#{token}"
The edge must produce the identical key or the two layers count separately.

note conductor

Two accounts poll on a fixed 200ms interval with no jitter, so the sliding window fills at exactly the same instant on every edge node.

Codrift memory view listing decision, summary, snippet and note entries with a full-text search box — FTS5 query syntax, straight through: redis AND fallback NOT dashboard.

06 / keys

Three views, two panes, one focus ring.

The keyboard is the interface, not a shortcut layer bolted on top. Views switch on a number, focus moves on a direction, and panes answer to their own number. The whole map comes from ~/.codrift/keybindings.json, so you can rebind any of it. The window below is a working model. Press the keys.

codrift 1 Context 2 Diff 3 Tree focus: sidebar

initiatives

rate limits

initiative.md

orchestration.md

memory

atlas-api

claude

atlas-edge

claude

dashboard

codex

adyen cutover

flaky test triage

Press the keys, or click them in the legend. On this page they are all ⌃: in the desktop app the split and the pane jumps are ⌘D, ⌘⇧D and ⌘1, but your browser owns those here and we are not taking them.

the sidebar has focus

: move the cursor
: into the pane
s: start an agent
b: branch the directories

a pane has focus

: focus left / right
: focus up / down, stacked
: jump straight to a pane
: switch this pane's view

any time

: split side by side
: split stacked
: collapse the sidebar
: command palette

Why leaving a terminal takes a modifier

Focus a pane running an agent and ⇥ stops being ours. A shell wants it for completion, and Claude Code cycles its own modes with it, so the pane hands every keystroke to the PTY and the model above shows it landing there. That leaves ⌃← as the way back out. It used to be ⌃⎋, which broke for an unpleasant reason: release the modifier a moment early and a lone ⎋ reaches the agent, which reads it as interrupt the running command. Directions have no such half-pressed state.

07 / surface

Sixty-six themes, and the one you already use.

Codrift reads VS Code theme JSON. Sixty-six are bundled; drop any other into ~/.codrift/themes and it shows up in the picker. Theme and font apply live to the whole surface, terminals included, so the agent panes match everything else on your screen.

Codrift theme picker filtering sixty-six bundled VS Code themes — Live preview with ↑↓, ⏎ keeps, esc reverts.

Codrift font picker showing installed monospace families and a size stepper — Your installed monospace families, with a size stepper.

The same Codrift tree view rendered in the light Catppuccin Latte theme — Same window, Catppuccin Latte. Light works too.

Pull the work in, don’t retype it

Connect GitHub, GitHub Projects, GitLab, Linear or Linear Projects and import an issue straight into an initiative’s context folder, where every agent picks it up automatically through --add-dir.

Browse and edit without leaving

3 opens a gitignore-aware file tree with syntax-highlighted previews, a fuzzy filter on /, and a CodeMirror editor with vim bindings on e.

Fuzzy file filter in the Codrift tree view narrowing to three rate limit files — Typing ratlim finds lib/atlas/rate_limit.ex.

08 / install

One binary. No runtime to install.

A native desktop app for macOS and Linux. No Node, no Erlang, no Ruby on your machine because the runtime ships inside. Bring your own agent CLIs and Codrift drives them.

MIT licensed, source on GitHub, free to run.

the short way

Puts Codrift.app in Applications and the codrift CLI in ~/.local/bin. Or pick one below.

↓ macOS .dmg ↓ Linux .AppImage / .deb

Not yet code-signed or notarized. The cask strips the download quarantine; if macOS still objects, right-click → Open once. Then codrift mcp install registers the MCP server with Claude Code.

From nothing to four running agents

n New initiative. Name it after the work, not the repo.
a Add the directories that work touches. Three repos is a normal number.
b Put every git directory on its own branch.
s Start an agent in each. The sidebar tells you which one needs you.

That is the whole loop. Everything else on this page is what happens after step four.

codrift update reads the releases API, checks the tarball against the SHA-256 published beside it, unpacks into a staging directory, and swaps it in with two renames, so a download that dies halfway never leaves you with half a release. It will not touch a Homebrew-managed install: rewriting a Cellar prefix behind brew corrupts its manifest, and the next brew upgrade would undo the whole thing. And when another codrift sits earlier on your PATH, it tells you, rather than updating a copy you are never going to run.

--check exits 0 when there is something newer and 1 when there isn’t, so it drops straight into a shell prompt or a login script.

Teach your agents to drive it

The MCP server hands agents the tools. The skills teach them when to reach for them: search the shared memory before exploring, check what the other agents in the initiative are touching, write findings back so the next session starts ahead.

codrift — the workspace model
codrift-memory — the shared store
codrift-initiatives — dirs, worktrees, diffs
codrift-orchestration — fan-out & conductor
codrift-integrations — GitHub, Linear, GitLab

Installs into Claude Code, Codex, Cursor, Gemini CLI, Copilot and Opencode. Add -g to install for every project. Or run Setup from the command palette (⌃P) and Codrift opens a terminal that does both this and codrift mcp install for you.