Terminal-native, no IDE lock-in
borg runs where you already work: your shell, over SSH, in CI. One command drives a full agent that reads, edits, runs and verifies. No editor extension to babysit.
ai coding agent · open models · honest pricing
The AI coding agent
macOS & Linux · runs as
The AI coding agent
that runs in your
terminal.
borg reads your codebase, writes the change, runs your tests, and fixes its own mistakes. An open alternative to Copilot, Cursor and Claude Code, on models we fine-tune ourselves. Start free, no card.
curl -fsSL https://turborg.com/install.sh | sh borg or turborg ·
open source on GitHub $ borg "fix the failing checkout test" · bash go test ./internal/checkout/... ✗· read_file internal/checkout/total.go ✓· edit_file internal/checkout/total.go ✓· bash go test ./internal/checkout/... ✓ ok internal/checkout 0.21s Fixed an off-by-one in the tax rounding. Test is green.
$ borg "add a /health endpoint that returns JSON status" · grep "func .*Router" ✓· write_file internal/api/health.go ✓· edit_file internal/api/router.go ✓· bash go build ./... ✓ Added GET /health returning {"status":"ok"} and wired it into the router.
$ borg "why does login take ~2s?" · grep "func Login" ✓· read_file internal/auth/login.go ✓· read_file internal/auth/hash.go ✓ The bcrypt cost is 14 (~1.6s per hash). Dropping to 12 gives ~0.2s and is still safe.
a real borg session · reads · edits · runs · verifies
An alternative to Copilot, Cursor and Claude Code.
The IDE assistants are great at finishing your line. borg does the whole job: it takes a task and works your repo end to end, in the terminal, on AI you are not locked into.
Your code stays yours
It works against your repo and sends only what a step needs. We do not train our models on your code, and you can run it on your own machine or your own shell.
Open models, pricing you can see
borg runs on open-weight models we fine-tune ourselves, billed by tokens, not per seat. Cheap enough to include in every plan, with no lock-in to a single vendor.
One command. The full agent loop.
- 01
Reads
Greps the tree, opens the files that matter and builds real context before it changes a line.
- 02
Edits
Writes the change across as many files as it takes: new files, refactors, fixes, in your actual repo.
- 03
Runs
Runs your real commands: tests, builds, linters, git. The same ones you run by hand.
- 04
Verifies
Reads the output, catches its own mistakes and keeps going until the work is actually done.
A coding agent that does real work.
Multi-file changes
Not autocomplete. borg plans a change, edits every file it touches and keeps the repo consistent as it goes.
Runs your real toolchain
Tests, builds, linters, formatters, git, package managers. If it runs in your shell, borg can run it and read the result.
A self-checking loop
It does not stop at a guess. borg reads command output, retries on failure and only reports back when the task passes.
Any language, any repo
Go, Python, TypeScript, Rust, whatever you ship. borg works at the shell, so it is not tied to one ecosystem or editor.
Skills and MCP
Teach borg repeatable skills and connect MCP tools, so it fits the way your team already builds instead of the other way around.
Runs on your shell
Install it on your laptop, or run it on an xShellz shell over SSH so the same always-on agent is there from any machine.
Open models
borg runs on models we fine-tune ourselves.
We build on the best open-weight models (DeepSeek, Gemma), fine-tune them on our own data, quantize them to run lean and wrap them in an agent harness. We do not pretrain from scratch, and we are honest about it. The payoff is yours: AI that is fast, cheap enough to include in every plan, and not locked to a single vendor.
Inside our AI research →- open weights
- fine-tune
- own dataset
- quantize
- distill
- agent harness
- evals
- fast inference
Beyond the terminal: turborg
The same agent, always-on across your networks.
borg works a session and exits. turborg is its always-on sibling: it holds your connection around the clock and layers the same AI on top, with message polish, channel summaries, agents and skills. The IRC connector is live and battle-tested across 17 networks, with more connectors coming.
Explore turborg → irc.libera.chatirc.oftc.netirc.dal.netirc.undernet.orgirc.efnet.nlssl.ircnet.ioirc.quakenet.orgirc.esper.netirc.hackint.orgirc.tilde.chatirc.alphachat.netirc.lizardirc.org
turborg · IRC connector · AI Polish · channel summaries
People already run their networks on turborg.
I'm mainly using it as an IRC bouncer and I've been very satisfied. The server's been stable and the control panel is responsive and easy to use.
The one-click summary feature is a lifesaver :) I leave my channels open but never actually read them. Now I just click the button when I wake up and get the TL;DR. UI looks nice too, dark mode is awesome! 5 stars.
Finally an IRC client that doesn't look like it was built in 1995 but also doesn't lag like Discord. It's good.
Token-based, not per seat. Start free.
One account covers borg and turborg. You pay for the AI you use, on a plan you can read top to bottom.
Free
$ 0 /mo
Start freeStarter
$ 10 /mo
See planPro
$ 24 /mo
See planMax
$ 60 /mo
See planFull breakdown of tokens, limits and features on the pricing page →
New AI, on infrastructure we have run since 2009.
xShellz is not a weekend project. The shells, runtimes and networks under borg and turborg have been online for over fifteen years, and the whole thing is open source. The intelligence is new. The ground it stands on is not.
180k+ Developers and members
15+ Years online, since 2009
OSS Open source on GitHub
24/7 Always-on, never sleeps
Frequently asked questions
Can't find your answer? Ask support →
What is borg?
borg is an AI coding agent that runs in your terminal. You give it a task in plain language and it reads your codebase, makes the change across as many files as needed, runs your tests and commands, and checks its own work before reporting back. It is the agent, not an autocomplete plugin.
How is borg different from Copilot, Cursor or Claude Code?
Copilot and Cursor live inside an editor and shine at inline suggestions. borg is a terminal-native agent in the same family as Claude Code: it drives your real toolchain end to end. The differences that matter for us are that borg runs on open models we fine-tune ourselves, bills by tokens you can see instead of per seat, and does not train on your code.
Which models does borg run on?
Open-weight models (built on DeepSeek and Gemma) that we fine-tune, quantize and wrap in an agent harness ourselves. We do not pretrain from scratch and we are honest about that. The goal is AI that is fast, efficient and cheap enough to include in every plan.
Does my code leave my machine?
borg only sends the context a step actually needs to the model, and we do not train on your code. You can install it locally, or run it on an xShellz shell you control. Either way the repo stays yours.
Is it really free to start?
Yes. The free plan needs no card and includes a real token budget so you can put borg on an actual task. When you outgrow it, Starter, Pro and Max raise the limits. You only pay when you outgrow the free tier.
Is borg open source?
Yes. borg and its turborg sibling are developed in the open on GitHub and we contribute back constantly. xShellz runs a managed, always-on instance on infrastructure we have operated since 2009, so there is nothing to deploy or babysit.
What is turborg?
turborg is the same AI, always-on beyond the terminal. It holds your connection to the networks you live in (IRC today, more connectors coming) and adds message polish, channel summaries and agents on top. One account covers both borg and turborg.
Put borg on your next task, free.
No card, no seat fee. Install it or sign up and run a real job in minutes.