Everest by Acific LabsEverest
by Acific Labs
The relay between models.
Version 0.10.1 · Runs on your own machine · This page: 0 trackers · 0 cookies · 0 third-party requests
What it is
Change the model.
Keep the project.
An AI agent is a program that does work for you, such as writing code. You might use one model to build, another to fix, and another to deploy. Everest runs beside them on your own computer. It relays what was done, what you corrected, and what was deployed, so the next model continues the same project instead of starting over.
An example
Three models.
One project.
You do not start over when you change models. The next one is given the project.
GrokBuilds it
The work is written down.
You ask one model to build the app. Everest keeps what was asked and what that model did, on this project, on your machine.
YouCorrect it
"Use the blue button."
You say what was wrong, in your words. That correction stays on the project. It is shown again to every model that comes after.
ClaudeContinues it
It does not start over.
You switch models. The new one receives the work so far, your corrections, and anything that was already deployed. Then it continues.
How it works
One log.
Every model appends.
The model is a utility. The project is the work. Everest is the log they share.
01Note
Each model writes what it did.
Point your apps at everest proxy and send the header x-everest-project. Every reply on that project is appended to the same log.
02Handoff
A different model receives it.
When the model changes, Everest puts the recent work in front of the new one. The same model, continuing its own thread, is not shown that thread again.
03Correction
Your words travel with the project.
A correction is kept in the words you used. It is included in every handoff, including when the same model continues. It is a record, not an order.
04Deploy
What shipped stays on the record.
When a reply says something was deployed, or you record it yourself, that line stays in the handoff. The next model can see what is already live.
AlsoBy hand
everest relay prints the handoff.
Run it in the project directory. Copy what it prints into the next model. --note records a correction or a deploy the proxy did not see.
AlsoAny model
One front door, several servers.
EVEREST_ROUTES sends each model name to its own server. Claude Code, Cursor, Codex, Gemini CLI and any OpenAI- or Anthropic-compatible app can share the project.
OvernightLessons
What repeats is still learned.
The relay is the handoff for this project. Overnight, Everest still turns repeated surprises into short lessons for the agents on this machine. A lesson describes what was seen. It is never an instruction. Delete its file and it is never written again.
Everest needs no model of its own. With one on your computer, lessons are written in better English and matched by meaning: everest model setup picks a model that fits your computer's memory and downloads it through Ollama.
The words
Six words
for the relay.
Project
One piece of work, named by its directory or by x-everest-project. Every model that touches it shares one log.
Model
Whoever is doing the work right now: Grok, Claude, GPT, Gemini, or a model on your own machine. The model can change. The project does not.
Handoff
The text the next model is given: the recent work, your corrections, and what was deployed. Observations, not instructions.
Correction
Something you said was wrong, kept in your words, and shown on every later handoff.
Deploy
A line that says this project was shipped. It stays on the record so the next model does not ship it again by guesswork.
Relay
The log itself, one file per project at ~/.everest/relay/. everest relay prints it. The proxy and the MCP tool relay write it.
What the next model sees
The handoff
is a record.
This is the shape of everest relay. It names the project, who has already worked on it, what they did, what you corrected, and what was deployed.
Project: harbor Models so far: Grok, Claude Work so far: - Person: Build the sign-in screen. - Grok: Added the screen and a blue primary button. - Person: Use the blue button, not the red one. - Claude: The primary action is the blue button. Corrections, in the person's words: - "Use the blue button, not the red one." Deployed: - [Claude] The sign-in screen is live at harbor.example.
A different model receives the work so far. The same model, whose own thread already has the last reply, receives the corrections and the deploys. Nothing in the handoff tells the model what to do.
Go further
See it, read it,
build on it.
The code is private until launch. Write to us if you would like to try it.
See it
The project, today.
The live page opens with the handoff, then the real memory on our machine. The handoff example is the shape, not that machine's log.
Use it
The user guide.
How to relay one project across models, then how the nightly lessons work. Every command, with what it prints.
Understand it
The paper.
Why agents need this, how it is built, what we have measured, and where it could go wrong. For sharing between machines, see the protocol in the menu.
What you can count on
The project.
On your machine.
- It stays with youThe relay log is a file on your computer, one per project. Nothing in it is sent anywhere unless you turn sharing on, and sharing moves lessons, never the project log and never your conversations.
- It never gives ordersA handoff and a lesson describe what happened. They do not tell the next model what to do.
- You have the last wordRead the handoff with
everest relay. Record a correction or a deploy with--note. Delete a lesson's file and it is never written again. - Where it standsVersion 0.10.1. The relay is the product: one project, any number of models. The memory around it still runs. It is early, and it works today.
- The nameEver, as in kept. It also sounds like the mountain: a rope left for the next climb. The handoff is that rope, from one model to the next.