Why Althar
The best coding agent right now is Claude Code. Or Codex. Or an open model. Ask again next month. New models ship most weeks, open models keep closing the gap, and prices and limits move with all of it. And when you swap, all your chats, your context, your integrations, lost.
The industry pace rewards optionality. The tools we use punish it.
Meet Althar:
- Everything stays on your computer. Projects, threads and history live in a local database on your machine. No Althar account, no Althar cloud, nothing of ours between you and the model.
- Open source, for real. Apache 2.0. Read every line, fork it, run it. If we disappear, your setup doesn't.
- Every plan, every account. Claude, ChatGPT and your API keys, side by side, with as many accounts on each as you have. Yes, we know, the second account is worth it. Althar uses them all.
- Runs out? It swaps. When an account hits its limit, the work moves to the next account or agent with room: same thread, same plan, same branch. No more waiting for a reset.
- Reviews, baked in. Ever felt like a meat proxy, pasting one agent's answer into another and typing "review this"? Every task is reviewed by another model, and the lead fixes what it finds before you see it.
- Works with your tools. GitHub, GitLab, Bitbucket, Linear, Jira and Trello. Issues in, pull requests out, named the way your team names them.
Subscriptions come and go. Your projects stay. Althar, stays.
What it does
And the rest
- Projects of one repository, several, or a folder inside one, in tabs across the top, with a home across all of them.
- Talk to any lead, any time. Interrupt it, queue a message for its next turn, change its model, or hand the task to another agent.
- Changes file by file before you accept them. Push stays yours.
- Code hosts and trackers: GitHub, GitLab and Bitbucket for pull requests; Linear, Jira, Trello and GitHub issues for work coming in.
- Never miss a call. A notification when something needs you, and a count on the Dock.
- Dictation that stays on your computer. After a one-time model download, nothing you say leaves it.
How it works
Everything runs on your machine. The project's tasks, threads and history are kept in a local database you can read.
What agents may do on your computer
Each task works in a git worktree of its own, under ~/Althar, never in your checkout. What an agent asks to do beyond its sandbox comes to Althar, which answers by the project's rules (Rules, in each project):
- By default, what stays in the task goes ahead. These wait for you: pushing to the default branch, force pushes, pushing every branch or tags, deleting branches that aren't the task's, deploying and publishing, and writing outside the task's worktree, installing software for the whole computer among it.
- If the coordinator decides, it answers the rest for you, never what waits for you or what could lose work.
- Allow everything, for a project or one task, lets everything go ahead but what's under Never and what could lose work. These still wait for you: a force push, pushing to or deleting the default branch, deleting or moving by force a branch that isn't the task's, dropping or clearing git's stash, git throwing away work in another checkout or replacing a file there, deleting anything outside the task's worktree, a delete whose target Althar can't read until it runs (
xargs rm, rm -rf $DIR), and reading a .env file outside it. Always allowing writes outside the worktree doesn't let these through either.
- On Windows, the same holds for cmd's and PowerShell's commands, such as
rd /s, del and Remove-Item -Recurse, and for paths with drive letters.
- Never, in any project: reading your keys and sign-ins (
~/.ssh, ~/.aws, ~/.npmrc, the keychain and the like), or changing anything on GitHub or GitLab except through Althar. Agents run without your SSH agent, git's credential helpers and gh's sign-in.
Althar's own git never force pushes unless you press Push on rewritten commits, merges into your checkout only as a fast-forward with nothing uncommitted there, and never deletes your branches or worktrees.
The rules see what reaches Althar. A command an agent's sandbox runs without asking (Claude Code and Codex run what stays in the worktree), and what a script does beyond its words, they don't. Keep work you care about committed and pushed.
Your agents
Althar runs Claude Code and Codex through the ACP adapters it ships, and the others as you installed them. It starts each in a mode where it asks, and answers from your project's rules; what each still does without asking is in its registry entry.
Get started
[!NOTE]
Early. Althar runs on macOS, Windows and Linux, and is tested most on macOS. Things will change, and some will break. Tell us when they do, or, in the app, Report a problem (in Help on a Mac, or at the foot of the quick settings), which sends what happened with the app's logs, paths and repository names hidden. Crash reports go on their own, masked the same way and never with your code, until you turn them off on the first screen or in Settings.
Sign in each agent you want to use with its own tool first:
claude auth login
codex login
opencode auth login
The others sign in with their own tools too: copilot login, junie, grok login, kimi login, vibe --setup, and qwen then /auth.
Then download Althar for your system, open it, and pick the repositories your first project works in. The first screen shows which agents are ready.
To build it from source instead, see DEVELOPMENT.md.
The thesis
Althar comes out of a broader hypothesis: what happens to software engineering once coding agents are abundant, and the project, not the agent session, is the thing that lasts. The argument, the questions it raises and the evidence we hope to collect are in THESIS.md.
Learn more
| |
|---|
| Development | Set up, run, test and change the code |
| Engineering standards | What every app and package in the repository holds to |
| Architecture | How Althar is built: a local-first desktop app, with seams for a later cloud |
| Decisions | What has been decided, and why |
| Open questions | What hasn't been decided yet |
| Glossary | The words the interface uses |
| Components | The interface, piece by piece, with every state as a story |
| Brand | The mark, lockups, banners, screenshots and the crew |
Contributing
The most useful thing you can do right now is use Althar on real work and tell us what breaks, in an issue. To change the code, start with CONTRIBUTING.md. Everyone taking part follows the code of conduct. To report a security problem, see SECURITY.md.
Licence
Althar is open source under the Apache License 2.0. See NOTICE.
Claude, Codex, Gemini, OpenCode, GitHub, Copilot, JetBrains, Junie, Grok, Kimi, Mistral, Qwen, Cursor, GitLab, Bitbucket, Linear, Jira and Trello are trademarks of their owners. Althar isn't affiliated with or endorsed by them.
