Ada
> < v0.1.60 · Windows, macOS, Linux
Ada is a desktop editor built around a coding agent that maps your repository once, sends only what a turn actually needs, and routes to any of 340+ models. Every change lands in its own git worktree until you merge it.
> < 01 / what one request carries
Both agents were asked to summarise the same project, on the same machine, from matched git checkouts. This is what each one put on the wire.
3× less sent. Ada's entire request is smaller than the part of the other request that exists before any work has happened.
> < 02 / how
No permanent system prompt riding along on every turn. Ada assembles each request from scratch, so a short question costs what a short question should — “hi” is about 272 tokens, tools and repo map included, because neither is sent.
Every tool you expose is re-sent on every request. An ordinary coding turn carries 1,772 tokens of schemas instead of 3,848 — decks, documents, images, notebooks, the browser and the design corpus stay out until the conversation asks, each behind its own trigger.
Opening a project builds a compact map of files and symbols, and the semantic index runs a local embedding model — no embedding API bill, ever, and your code never leaves the machine to be indexed.
Asked for a presentation, Ada calls its deck writer and hands back a real
.pptx: $0.28 against $1.25 for an agent hand-writing slide markup.
Same for lookups — the map answers without opening files.
> < 03 / the editor
Repositories sidebar, Monaco, fuzzy open, find-in-files, an integrated terminal, a review pane. No learning curve. What's different is that the agent is the main pane — and several of them can run at once.
Ctrl+P for files, Ctrl+Shift+F for text across the project.
No global run slot. Each chat streams with its own queue and parked questions; the list shows which are working and which are waiting on you. Background jobs survive an agent restart and report into a docked tasks panel.
One toggle chooses the model tier and the shape of each turn — a normal tool loop, a plan-first pass, or a fan-out over a file too big for the window, where one worker per chunk reads its slice and nothing is summarised before the question is asked.
Sign in with a Claude or ChatGPT plan and that vendor's models bill to the plan, served by a local gateway — the token never leaves the machine. Ada can also drive the Chrome you already have open, logged-in sessions included, through a small bridge extension.
Any prompt can run on a schedule. Marketing turns that into campaigns: scan a product, get plays with a cadence, edit them in place, approve each draft before it posts. Usage is a bar and a percentage — dollars per four-hour window, never a figure on screen.
> < 04 / the benchmark
Ada against Claude Code — same machine, same model (Opus 4.7), same prompts, matched git worktrees at the same commit. Every run, including the one Ada lost.
| Task | Ada in | Ada cost | Other in | Other cost |
|---|---|---|---|---|
| Summarise the project | 31,856 | $0.2046 | 95,941 | $0.5291 |
| Create a pricing API | 206,954 | $1.2340 | 424,706 | $1.5339 |
| Tetris game | 109,041 | $0.7634 | 95,708 | $0.6885 |
| Portfolio website | 57,075 | $0.5674 | 128,389 | $0.6738 |
| Project presentation | 38,726 | $0.2807 | 148,150 | $1.2517 |
| Total | 443,652 | $3.0501 | 892,894 | $4.6770 |
Read the Tetris row closely. It's the one task that cost more — by seven cents — and it shipped 6.7× the code: 630 lines against 94, with a next-piece preview, levels, a ghost piece, pause and wall-kick rotation the cheaper build doesn't have. Per hundred lines delivered that's $0.12 against $0.73. Tasks anchored in existing code ran 16–78% cheaper, and the gap widens as a codebase grows.
Read the full report — screenshots, method, limits →> < 05 / your own providers
Ada talks to any OpenAI-compatible endpoint. OmniRoute is one you run on your own machine: it holds your provider accounts, and Ada treats it like any other backend. Settings → Connections installs it and connects, in one click.
Ada installs it into a folder it owns, starts it, and points itself at it — no
sudo, no npm prefix to reconfigure. If Node is too old or the port is
taken, the card says which, and what to run.
Providers are connected in OmniRoute's own dashboard, on localhost. Ada
never sees those keys, and nothing you send through the gateway is metered or billed
here.
Disconnect points Ada back at its own backend, and the gateway keeps running — it is yours, not Ada's. Removing it is deleting one folder.
OmniRoute is a separate open-source project, and you bring your own accounts to it. It needs Node 22.22.2 or newer — Ada checks before installing rather than after. OmniRoute on GitHub.
> < 06 / download
Install and chat immediately with the free models — no account. Sign in with GitHub to unlock the full catalogue.
macOS builds are signed and notarized by Apple. On Windows, if SmartScreen appears, choose More info → Run anyway. Ada updates itself from then on. All versions.