EVΛL
The safest way to let
an agent touch production.
Coding agents write faster than anyone can review. EVΛL runs them on your own server with a permission tier enforced by the tool list, a gate they cannot bypass, and an owner who approves anything irreversible from his phone.
What it runs
The plan, the agent, or your own commands.
Every session is the same thing: a project directory, a permission tier, and a transcript. What you point it at is up to you.
The plan
In both mode GPT holds the conversation and writes the plan; Claude Code executes it and runs the tests. You read one thread.
# what GPT hands over
## Goal
Make the composer show who is executing.
## Acceptance criteria
- Both targets compile.The agent
Claude Code and Codex already live on the server, inside the project, with the same files and rules a human engineer reads.
slink-claude "fix the failing STT check"
# sessions resume; jobs keep
# running after you close the appYour own commands
Builds, migrations, container restarts. One tap from the phone, with the output streamed back as it happens.
$ docker logs --tail 200 slink-api
$ ops/stage.sh
358 passed · 57/57 checksIsolation
Avoid unintended access
to your live repository, your secrets, and the services your customers depend on.
A permission tier is a different tool list, not a different instruction. Under plan the write tools are not present in the session at all, so there is nothing for a model to disobey — and nothing for a prompt injected into a file to call.
The gate
Protect against a confident wrong answer
Agents are most dangerous when they are sure. Every change EVΛL makes passes the same four steps, in order, with no way around them.
$ ops/promote.sh
refusing: merge would delete 478 files
# the guard ran before the merge,
# not after it
$ ops/promote.sh
promoted 060bc3a
Live is healthy.That first line is real. A test of ours once wrote into the shared git index and produced a commit holding one file; the guard refused the merge three times before we found the cause.
Permissions
A tier is a tool list, not an instruction.
This is the whole idea, so you should be able to see it rather than take our word for it. Pick a tier and watch the session's tool list change. Under plan the write tools are not refused — they are not there.
Questions
Frequently asked
What is EVΛL?
A development environment that runs coding agents — Claude Code and Codex — on your own server, inside your project, driven from a phone or a desktop. It is the environment that builds and maintains SLINK itself.
Where does the code run?
On your server, in the project directory you chose. Nothing is uploaded to a third-party execution service. Sessions resume, and a job keeps running after you close the app.
How is a permission tier enforced?
By the tool list handed to the model for that session. Under plan the write and shell tools are not present, so there is nothing to disobey. Under ask the risky calls stop and request approval on the owner's phone. Under full every tool is pre-approved, including the network and the server.
Can an agent deploy to production?
Only through the gate: unit tests, a staging deploy, 57 live checks, then promotion with automatic rollback. The live tree is written by ops/promote.sh alone — no agent commits into it directly.
What does “both” actually do?
GPT leads the conversation, because it is stronger in Arabic dialogue, and writes the plan in English Markdown — Goal, Steps, Acceptance criteria. It hands the plan to Claude Code, which edits the files and runs the tests, then reports back in Arabic. Two models, one turn, one thread to read.
Which models can I use?
Claude Opus, Sonnet and Haiku; GPT and Codex from OpenAI; and 382 more through one OpenRouter key for agency roles. Batch-only variants are filtered out, because they cannot answer inside a live loop.
Is EVΛL for sale?
Not today. It is the founder edition and is not compiled into the commercial build. The capability behind it — an agency of models with roles and enforced permissions — is what we are bringing to the SLINK app.
Start