Software that fixes itself.
Open source and self-hosted. Pre-alpha Open app
It lives on your box, watches production and hunts bugs.
Each one becomes a quest: evidence, repro, patch.
It proves each fix. You approve what matters.
It lands on Radicle, Tangled or plain git.
Good fixes set. Bad ones undo themselves.
Jellyware: soft enough to change, firm enough to trust.
A workspace for you and your agents.
Connect your application, follow the work, and bring your team into the decisions. Explore the actual Ephyra interface below.
Sample data. Open a screen to explore it. Changes stay in this preview; no repository, invitation, or model call is created.
Open your own workspaceOne bug, start to finish
A sample checkout bug, from the first error to the watched fix. Follow it in the graphical console or the terminal.
-
01 triage
It notices, and decides it’s worth a fix
Thirty-seven checkout errors share one fingerprint. Two small models triage them: real, not a duplicate, medium risk. A quest opens in git, and an agent starts on it.
-
02 checkpoint A
Is this the right cause?
The agent writes down what it thinks is wrong, and why. You read a page, not a diff. In checkout, policy says a person agrees before any code is written: press a.
-
03 evidence
Ask why
Press e to see what the hypothesis rests on: the error stream, the commit that started it, past quests, and the agent session that wrote that commit.
-
04 checkpoint A
Or push back
Press r and write a note. The agent gets it with the quest and tries again. Here, nothing is built until someone approves.
-
05 checkpoint B
Reproduced, and a plan
It reproduced the bug with a probe that fails on main, then planned the smallest fix. Comments stay on the quest, where the agent and the next reviewer see them.
-
06 checkpoint C
Proof, before you look
The fix arrives with its evidence. The probe passes now, so do 412 tests and six regression probes, replayed traffic shows nothing unexplained, and the diff is two files. Approving lands it.
-
07 landing
It lands where you work
This one lands as a Radicle patch. It could be a Tangled pull, a plain git ref or a live change in your devenv. Then it watches production for two hours.
-
08 watch
It sets
Two quiet hours later the quest closes. Its repro joins the regression probes, so this bug can’t come back unnoticed.
-
09 rollback
Or it undoes itself
A different fix, on a low-risk edge, landed without asking anyone. Twenty-seven minutes later a new error showed up in the file it touched, so Ephyra reverted it and reopened the quest.
You decide where it asks
One Markdown file in your repo says where a person decides, how a fix lands and how long it’s watched.
Checkout asks at A, B and C. Docs can land on their own once you enable that area. This example starts in shadow mode: every quest still waits for you.
# Autonomy policy ## defaults mode: shadow checkpoints: A=advisory B=advisory C=blocking landing: trail_ref watch: 2h ## area: checkout paths: services/checkout/** risk: medium checkpoints: A=blocking B=blocking C=blocking landing: radicle_patch watch: 2h Money moves here. A person reads the hypothesis and the plan before any code. ## area: docs paths: docs/**, **/*.md risk: low checkpoints: A=advisory B=advisory C=advisory landing: auto_land watch: 1h
The inbox, by key
- a
- approve the checkpoint; at C, this lands it
- r
- send it back with a note
- e
- explain: the evidence behind it
- t
- claim a checkpoint review; release it before deciding
- x
- kill the quest, with a reason
All the keys
Explore the pre-alpha
Ephyra is open source and pre-alpha. It is built to run on your own hardware or Railway, with no GitHub, and to land on Radicle, Tangled or plain git. The design explains how quests, checkpoints and landing work.