NewClinkX 1.0— multi-model orchestration, built inSee how ↗
The runtime for autonomous coding agents

Delegate the work. Trust the result.

Elaris gives every agent a full, isolated machine — editor, terminal, browser, live previews — with ClinkX orchestrating the models behind it. Hand off a task, watch the work happen live, ship the pull request that comes back.

launch by api · orchestrate via mcp · isolated by default
session el-7f2clive
taskadd retry with backoff to the api client
readsrc/providers/client.ts
editclient.ts +41 −6
runpnpm test --filter providers
pass18 tests passed in 6.2s
openpreparing pull request…
agent working · 2 files changed:3000 live
FIG. 001 — LIVE AGENT SESSION
00:00 — task handed off14:32 — pull request opened

One session, every model — ClinkX routes each stage to the model that does it best

ClaudeGPTGeminiZ.AIOpenRouter
The work, in the open

A machine of its own, watched live.

session el-7f2c · live
// launched via api — no human at the keyboardPOST /v1/agent-sessions  202 acceptedsandbox up in 1.8s · repo clonedediting src/providers/client.ts18 tests passed · preview on :3000
FIG. 002 — SESSION LAUNCH
How it works

From prompt to pull request, in three moves

elaris · api
POST /v1/agent-sessions202 acceptedsandbox up · repo clonedports ready · session live
FIG. 003·A — SESSION LAUNCH
agent · activitywatching

readsrc/providers/client.ts

editclient.ts +41 −6

runpnpm test --filter providers

pass18 tests passed in 6.2s

servepreview forwarded on :3000

FIG. 003·B — AGENT AT WORK
src/providers/client.ts · diff+41 −6
async request(url, opts) {−  return fetch(url, opts)+  return withRetry(() => fetch(url, opts), {+    attempts: 3, backoff: "expo",+  })}
18 tests passedopen pr ↗
FIG. 003·C — REVIEW & MERGE
01

Launch a session

One API call — or one ClinkX tool call from any MCP client — and Elaris spins up an isolated sandbox with your repo, toolchain, and ports ready.

elaris · api
POST /v1/agent-sessions202 acceptedsandbox up · repo clonedports ready · session live
FIG. 003·A — SESSION LAUNCH
02

The agent does the work

It edits, runs, and tests inside the sandbox — editor, terminal, and forwarded ports on one surface you can watch live.

agent · activitywatching

readsrc/providers/client.ts

editclient.ts +41 −6

runpnpm test --filter providers

pass18 tests passed in 6.2s

servepreview forwarded on :3000

FIG. 003·B — AGENT AT WORK
03

Review and ship

Every change is reviewable inline. Fork the chat, restore any checkpoint, then open the PR when the tests are green.

src/providers/client.ts · diff+41 −6
async request(url, opts) {−  return fetch(url, opts)+  return withRetry(() => fetch(url, opts), {+    attempts: 3, backoff: "expo",+  })}
18 tests passedopen pr ↗
FIG. 003·C — REVIEW & MERGE
No. 01 — The live workspace

Watch the work as it happens

Every session gets a full workspace you can open at any moment — the agent's editor, terminal, and live previews on one surface. Devtools traces every request and tool call, so you can replay exactly how the agent got there.

live session viewdevtools traceinline diffsterminallive previews
elaris · workspacesandbox: up
src
·ChatPage.tsx
·composer.ts
·sandbox.py
·README.md
export function Workspace() {  const sandbox = await spawn()  sandbox.run("npm dev")  // preview forwarded on :3000}
agent: editing · 2 files changed:3000 live
FIG. 004 — THE LIVE WORKSPACE
No. 02 — Orchestration

One orchestrator, every model

ClinkX breaks a job into stages and hands each one to the model that does it best — then validates the work and merges it into a single, coherent change.

multi-modelcontext bundlesstaged workflowsauto-merge
clinkx · orchestratormulti-model
run_workflow+ screenshot.png
ClinkX
Geminiperceive
Codexplan
Claudebuild
context built · delegated · validated · merged
FIG. 005 — ORCHESTRATION GRAPH
No. 03 — The evidence

Orchestration resolves more.

One model is a generalist. ClinkX splits a task across specialists — one to read, one to plan, one to build — then validates and merges. On every task type we measure, orchestrated sessions resolve more of the work end to end.

1.8smedian sandbox cold-start
6+models orchestrated per session
1:1isolated sandbox per session
FIG. 006resolution rate · %
1007550250
7194
Bug fix
5888
Feature
4982
Refactor
4179
Migration
single modelclinkx orchestration
FIG. 006 — RESOLUTION RATE BY TASK TYPE, SINGLE MODEL VS. CLINKX
No. 04 — Extensibility

Everything extends

Define custom agents and skills, wire up your own MCP servers, and add slash commands. Elaris bends to the way your team already works.

skillsagentsslash commandsmcp servers
extensions.json6 installed
agentreviewertools: read, grep
skillpdf-export.zipyaml + assets
command/reviewbuilt-in
command/compactbuilt-in
mcppostgresuvx · http
mcpplaywrightnpx
FIG. 007 — EXTENSION MANIFEST
No. 05 — Isolation

Sandboxed and isolated, by default

Each session works in its own private sandbox. Secrets are encrypted, access is scoped per session, and your code is never used to train anyone's model.

one sandbox per sessionencrypted secretsscoped accessprivate by default
encryptedscoped access
isolated runtime
private sandbox
FIG. 008 — ISOLATION MODEL
FAQ

Common questions

Our orchestrator and MCP server. It breaks a job into stages, hands each one to the model best suited to it, then validates and merges the result into a single coherent change.

Get started

Put an agent on it.

That fix you keep postponing, that refactor nobody volunteers for — hand it off and see what comes back.

01 connect a provider02 hand off a task03 review the pr