Spark
Answer, or one small edit
A question about your code, or a change small enough for one direct pass: a line, a value, a comment. No plan, just the answer or the edit.
Free edition · v0.2 · runs on your machine
Autonomous software engineering, under your control.
AutoSE turns a request in plain English into reviewed, checked code. It sizes up each request, plans the work, does it, and checks it against numbered requirements, and it asks you before it runs anything. All from a dashboard in your browser, on models you run yourself. Nothing leaves your machine.
Install, free
curl -fsSL https://autose.dev/install.sh | sh
irm https://autose.dev/install.ps1 | iex
Click to select
Installs AutoSE, adds a desktop shortcut and opens the dashboard. Needs git; the installer sets up uv and Python for you. Bring a local model from Ollama, LM Studio, vLLM or llama.cpp.
Try it
Run a session, right here.
Pick a request and watch a session play out the way it does in the dashboard. It all happens in this page: no model behind it, nothing installed.
Enter starts · Shift+Enter for a new line
How it works
Every request takes the route it needs.
AutoSE sizes up each request first, so a quick question stays quick and real work gets a plan and an independent check. When a check fails, the work goes back for another pass before it ever reaches you. The dashboard draws every run this way, with the time each step took.
-
Start Classify
Auto reads the request and picks a tier: Spark for a question or a small edit, Craft for anything that needs a plan. You can pick one yourself instead.
-
Forge Design Pro
With Forge, a bigger change is decided against your ADRs and standards first, and the decision is written down as a new ADR.
-
Step 1 Plan
An ordered plan with numbered checks that every later step is held to. Read it before any work starts, if you like.
-
Step 2 Execute
The plan is carried out in your project. Every command waits for your answer, unless you have allowed it.
-
Step 3 Validate
A separate step that cannot write files checks each requirement, with evidence. If one fails, the work goes back to Execute.
-
Then you Your review
See the diff of every file it changed, then keep it or undo the run.
Tiers
The right amount of effort for every request.
Not every change needs a plan, let alone an architecture review. Auto picks a tier for each request, or you choose one yourself. Every session shows its tier as an icon, so you can tell them apart at a glance.
Craft
Plan → Execute → Validate
Everything that needs a plan, from a bug fix to a change across many files. Planned, done and checked before you see it.
Forge Pro
Design, then as Craft
Architecture-level work gets a Design step first: decided against your ADRs and standards, and recorded as a new ADR before it is planned and checked.
Forge is part of AutoSE Pro. In the free edition, bigger changes run as Craft.
Features
Engineering you can account for.
Everything happens in a dashboard in your browser, in a light theme or a Catppuccin Mocha dark one, or in the terminal if you prefer. You decide what runs, and you can see what changed.
Local
Runs on what you own
AutoSE finds Ollama, LM Studio, vLLM or llama.cpp on your machine by itself. By default nothing leaves it: no third-party service, no telemetry.
Models
Made for small local models
It plans for the context window your runtime really serves, and tells you when that is smaller than the model allows. Small models sized to the task use far less energy than a giant cloud model.
Sessions
Several sessions at once
Run a long change in one session and ask a question in another. Rename them as you like; each shows its tier, and they pick up again after you close the window.
Route
See every step
Each run is shown as a route: every step with its timing, and every send-back marked where it happened and why.
Checks
Numbered checks
Requirements are numbered when the plan is written. You see each one go from written to done to verified, with the evidence for it.
Needs you
One list of what waits for you
Needs you gathers every question and plan waiting for an answer, from every session, so nothing sits unseen in a tab you are not looking at.
Questions
Asks before it runs anything
Commands wait in place with Run it, Always allow or Don't run, or tell it what to do instead. Always allow covers that kind of command, not your whole shell, and a chain of commands is judged part by part.
Allowed
See what runs without asking
The Allowed page lists every command that no longer asks first, so you can read it, change it or take it back.
YOLO mode
Nothing asks, when you mean it
One switch lets everything run without asking. A red frame stays round the window for as long as it is on, so you never forget.
Changes
Diff, undo and files
Read the diff of every change, undo a run, and browse the whole project in a read-only Files view.
Schedules
Schedules and watchers
Run a request on a schedule, or have a watcher run a command and start a request when its output changes or it starts failing.
Plan review
Read the plan first
Turn on Review the plan and the work stops once the plan is written. Approve and start, edit the plan, or stop there.
Benchmark
Long-horizon tasks on a local model.
We ran the free edition of AutoSE, mini-swe-agent and OpenHands on 20 DeepSWE tasks: whole features in active open-source projects, graded by hidden tests on the work the agent commits. Every agent got the same model, the same instruction and the same three hours per task.
AutoSE 0.2.5
OpenHands
mini-swe-agent
| Agent | Target tests passed | Work committed | Median time per task |
|---|---|---|---|
| AutoSE 0.2.5 free edition | 297/1020 29% | 12/20 | 160 min |
| mini-swe-agent 2.4.6 | 0/1020 0% | 1/20 | 156 min |
| OpenHands Agent SDK | 19/1020 1.9% | 1/20 | 180 min |
- Tasks. 20 of DeepSWE's 113 tasks (Go, Python, TypeScript, JavaScript and Rust), run with DeepSWE's own instruction and environments. The agent has no internet access; grading applies the agent's committed changes in a fresh container and runs the hidden tests, so uncommitted work scores nothing.
- Measures. Target tests are the fail-to-pass tests that check the requested feature, pooled over the 20 tasks. No agent passed every test of any task.
- Model. Qwen3.8-27B (Q4_K_M GGUF) served by llama.cpp, one replica per GPU, 128K context per request, thinking off. All three agents sample at temperature 0.7, Qwen's recommended setting.
- Harness. Harbor, DeepSWE's 3-hour limit and no step limit. AutoSE is the free edition at
main
a0febd2. Many runs reached the time limit: 9 for AutoSE, 10 for mini-swe-agent and 13 for OpenHands.
Editions
Free for every developer. Pro for teams.
The free edition is the whole agent: install it below and use it on any project. AutoSE Pro adds the decision-record and knowledge layer organisations need.
Free · on this site
AutoSE
SparkCraft
- The browser dashboard and the terminal, with several sessions at once
- Quick answers and small edits, and planned changes checked step by step
- Asks before every command, with Always allow and YOLO mode when you want them
- Diff and undo for every file, a Files view, schedules and watchers
- Sessions that pick up again after you close the window
- Finds Ollama, LM Studio, vLLM or llama.cpp on its own
Pro · for teams
AutoSE Pro
SparkCraftForge
- Everything in the free edition
- Forge: a Design step that records each architectural decision as an ADR
- Your ADRs and organisational standards brought into every run
- Citation checking, so no rule is quoted that was never shown
- Learned project facts, kept only once a person approves them
Get started
One line to your first session.
The installer works the same way on macOS, Linux and Windows. Run it again any time to update.
-
Install AutoSE
Run the one-line command for your system. It downloads AutoSE, sets up Python withuv, adds an AutoSE shortcut to your desktop and opens the dashboard. -
Set up a local model
AutoSE runs on a model you host. Install Ollama and pull a coding model that can call tools:ollama pull qwen2.5-coder:14b. AutoSE finds it by itself. -
Open it on your project
Double-click the AutoSE shortcut and choose your project folder; the dashboard opens in your browser. In a terminal,autose --webdoes the same for the current folder,autosestarts a terminal session andautose --doctorchecks your setup.
macOS / Linux
curl -fsSL https://autose.dev/install.sh | sh
Windows (PowerShell)
irm https://autose.dev/install.ps1 | iex
Want to read it first? Both scripts are plain text: install.sh and install.ps1. You can also build from source by following the README.