A native IDE where agents plan, build and review under your direction — not another chat window. Four surfaces over one workspace: talk to it, let it run, direct a crew, review every line.
The number nobody else shows you: what you saved.
Why does the refresh check get skipped on cold start?
Because skipRefreshCheck stays true while the daemon rehydrates.
The PTY daemon reattaches sessions before the auth layer mounts, so the first tick reads a stale flag. I traced it and have a clean fix —
⟶cargo test token_engine
Compiling octopush v0.4.50
Finished test [unoptimized] in 3.41s
Running tests/token_engine.rs
running 12 tests ............
test result: ok. 12 passed; 0 failed; 0 ignored
⟶▊
Local-first· No telemetry· Signed updates· Bring your own keys
Many models, one workspace — switch mid-session
Anthropic & OpenAI-compatible protocols with a custom base URL — point it at your own gateway, an OpenAI-compatible endpoint, or a local Ollama. No per-vendor code.
Direct assembles a pipeline of agents and gives each stage exactly the model it deserves — and exactly the reasoning effort it needs: a cheap model thinking deeply on synthesis, barely at all on the mechanical parts. A stage escalates to a stronger model only after a proven failure — you pay the expensive tier on a real failure, not by default. Watch one execute: the run track fills, the work journal records every tool call, and it pauses at your checkpoint.
This run saved $0.089 against the all-premium baseline.
Fork a template or wire your own: drag stages onto a node graph, set each one's role, model and substrate, and mark the gates where a human signs off. The same pipeline you build is the one that runs above.
API or CLI⌘⇧M) — every run across every workspace, one cockpitThe number nobody else shows you: what you saved.
Running five agents in parallel is a feature.
Giving each a role, a model, a budget and a human gate is a workflow.
Octopush refuses to flatten an editor and a chat box into one blurry pane. Terminal person, chat thinker, pipeline architect, careful reviewer — four deliberate surfaces over the same workspace, and it glides between them. Pick one to drive the demo above.
for thinking out loud
Streaming answers split into key phrase and body, tool calls as collapsible
cards, files mentioned with @, skills invoked with / —
and a $ prefix that runs shell commands in a persistent
per-thread shell, spending zero tokens.
for terminal natives
Genuine PTY shells via a persistent daemon. Many named terminals per workspace, clickable file paths into Review, and sessions that survive a restart.
for directing a crew
Design a multi-stage pipeline — a role, a model and a substrate per stage — with human checkpoints and a ledger that leads with what you saved.
for staying in control
A unified diff with per-hunk verdicts, an integrated editor, your own test command, and an AI review that points at the exact line it's worried about.
Octopush runs an out-of-process PTY daemon — a separate binary that owns every pseudo-terminal. Quit the app, install an update, relaunch: your shells are still running, scrollback intact, exactly where you left them.
18:42 — you quit the app
⟶npm run dev
VITE v6.0.3 ready in 412 ms
➜ Local: http://localhost:5173/
18:41:57 [vite] hmr update /src/App.tsx
18:42:10 [vite] hmr update /src/stores/runsStore.ts
18:47 — five minutes later
➜ Local: http://localhost:5173/
18:41:57 [vite] hmr update /src/App.tsx
18:42:10 [vite] hmr update /src/stores/runsStore.ts
18:46:52 [vite] hmr update /src/App.tsx
⟶▊
Agents write a lot of code. Octopush keeps you editor-in-chief: a unified diff with per-hunk accept and reject, an AI review that points at the exact line it's worried about — and a "Why?" trail that traces any edit back to the agent turn that produced it: which role, which model, which message.
Most tools hide the meter. Octopush puts input, output and cache tokens where you can see them — with per-model cost, today's spend, and budgets per workspace, project or globally that warn at 50, 80 and 100% before you blow past them. If you pay per token or bill per project, that spend is your cost of goods — so Octopush keeps it in the frame, not buried in a dashboard you never open.
Zero wasted tokens — measured, not promised.
There is no Octopush server reading your work. Projects, chats, run history and settings live in a SQLite file on your disk. Keys are stored locally and sent only to the providers you configure. There is no analytics SDK in the app to opt out of — none ships.
Turn sandboxing on and a mission's agents run write-confined to its workspace
by the macOS seatbelt — your ~/.ssh, ~/.aws, other worktrees
and system paths stay read-only, enforced by the kernel rather than asked of the agent.
Review missions are read-only by construction. And it's on the free tier:
security is never paywalled.
Everything lives in ~/Library/Application Support/octopush — a SQLite file you can open, back up, or delete.
API keys are stored on your machine and travel only to the provider you chose — Anthropic, OpenAI, DeepSeek, or nowhere at all with Ollama.
No analytics, no crash uploaders, no phone-home. Outbound traffic is exactly what you configured: your providers, your integrations.
Every update is Ed25519-signed and verified before install. What you download is what we built.
Built by a developer who runs it on his own production work, every day. Shipped weekly, versioned honestly, documented as it lands. Designed as an atelier — warm black, surgical brass, serif sentences — and the restraint to use them sparingly.
Tauri & Rust core. React 19 surface. A custom PTY daemon, a SQLite ledger and a provider router doing the quiet work.
Octopush ships its own MCP server: drive pipelines, projects and workspaces from Claude Code or any MCP client. It stages drafts — it never runs, spends, or mutates git on its own.
Read your backlog beside the work, open a workspace straight from a ticket, and see the open PR for every branch. Pick a GitHub issue and the "Ship it" crew folds it into the task and drives it to a PR.
⌘K to leap between workspaces, sessions, models and actions. ⌘1–9 jumps workspaces; ⌘⇧M opens Mission Control.
Stage, commit, push, stash, cherry-pick — through your login shell, so SSH agents, signing and hooks behave exactly like your terminal. AI drafts the commit message from the staged diff.
A quick code-and-text pane beside the chat — CodeMirror 6, a dozen languages, for the thought you don't want to lose.
Atelier by default, eight more built in — and a model catalog whose per-token prices refresh from live data, so the cost meter never lies.
ProDetached runs survive quitting the app and post a native notification when the crew finishes or needs you. Routines fire scheduled crews on a Days × Times schedule, with an optional shell pre-condition that skips a window when there's nothing to do.
No repo yet? Describe what you want to build and set a crew on it — the Greenfield pipeline is born into an empty project, scaffolds the first working slice, and you direct every gate. Sandboxed by default.
One cockpit (⌘⇧M) for every run across every workspace — needs-you, in-flight, settled. The Logbook keeps each mission's ledger: worked time, spend, and savings vs an all-premium baseline.
The documentation takes you from a cold download to running a pipeline — the mental model, every mode, the full shortcut sheet, and what to do when something misbehaves.
Octopush is free and local-first — bring your own keys and work the core without limits. Pro unlocks the multi-agent orchestration harness. Cancel anytime.
for everyday local work
for orchestrating a crew of agents
Download free, then unlock Pro in-app after you sign in.
Octopush is a native desktop app for macOS — Apple Silicon & Intel — with signed built-in updates. Bring your own keys: Claude, OpenAI or DeepSeek, or run a local model with Ollama. Free to start.