CloudCode · engine for Code Easy
v0.1 · beta

Multi-model coding cockpit

One conversation.
Many models.

A single terminal for driving many coding models — cloud, local, or your Claude subscription — in one thread. Switch mid-task, ask another model for a second opinion, compare two side by side, and run entirely on local models with no API key.

10 providers Ollama · LM Studio local no key · private MCP · skills · agents

The difference

The thread that survives the model switch.

Tab-switching between Claude Code and Codex means re-pasting context by hand every time. CloudCode keeps one shared conversation as you move between models — so "use the right model for this step" is one keystroke, not a copy-paste dance across windows.

Where it fits

Five things it's genuinely good at

Multi-model

One conversation, many models

Turn the model like a dial mid-task. Have one draft the architecture, hand the same context to a cheaper or local model for the boilerplate, then diff a tricky answer. Every reply is tagged with the model that wrote it.

/model/also/compare
Local-first

Local models as first-class citizens

Ollama and LM Studio run natively — no API key, nothing leaves your machine, no per-token bill, and answers stream token-by-token. Flip to your local model and back without losing the conversation.

/offline/localmodels
Compare · compress

Comparison and compression

Run two models on one prompt and read them side by side. Long sessions auto-summarize to fit smaller local context windows. Second opinions send only the last exchange by default — cheap on a long thread.

/compare/compact/multimodel
Trust

Trust tools for models you didn't train

The distinctive angle. Fingerprint and pin a local model file to catch a later swap; differential-test a prompt against a trusted reference; a runtime tripwire flags reverse shells, exfiltration, and secret-path access.

/integrity/crosscheck/sentinel
Extend

Extensible and embeddable

Connect any MCP server, add reusable skills, parallel sub-agents, project shell tools, tool-event hooks, and installable plugins. The same core is a library a host can drive headlessly.

.cloudcode/mcp.jsoncreateEngine()
Providers

Point it at anything

Anthropic, OpenAI, Gemini, DeepSeek, Qwen, Kimi, GLM, Ollama, LM Studio, or any OpenAI-compatible endpoint — plus your Claude subscription via the CLI bridge, no key.

anthropicollamalmstudiocustom/…claude-cli

Complementary, not a replacement

Orchestrates · executes · runs beside

CloudCode is additive. It's the engine Code Easy drives, and the panel you open next to Claude Code or Codex when you want a different model's take, a private run, or a safety check.

Orchestrates
Code Easy

Owns the spec → autonomous multi-agent build loop and the work queue. Selects a runner.

Executes
CloudCode

The interchangeable engine that actually writes the code — with provider choice, MCP, and the trust checks. Same runner contract as Codex.

Runs beside
Claude Code · Codex

Peers you keep using. /claude-cli even drives Claude Code on your subscription.

Two lenses

Read it your way

The same tool, framed for who's holding it — presented, fittingly, the way CloudCode presents two models: side by side.

For the engineer↳ the plumbing
  • Provider-agnostic core. One neutral message format; adding a provider is a single registry entry.
  • A real permission model. Deny beats allow and descends into command substitution; plan mode is a hard read-only gate; WebFetch is SSRF-safe at connect time.
  • Embeddable. createEngine() streams text/tool/cost events — exactly how Code Easy runs it.
  • Tested, not a black box. ~267 tests: a mock-provider loop, an Ink TUI harness, and dispatcher tests for every command.
  • Deploys anywhere Node 20 runs. A ~3 MB single file or a zero-dependency native binary.
For the vibe engineer↳ the feel
  • One place to talk to every AI. No juggling tabs. If an answer feels off, type /also for a second brain.
  • Free and private mode. Have Ollama or LM Studio? It just uses it — no key, no bill, nothing leaves your laptop.
  • It won't wreck your machine. Asks before anything risky; plan mode keeps it read-only while you explore.
  • It remembers and tidies up. Long chats get summarized so you don't hit limits; notes carry across sessions.
  • Use it beside your other tools. The panel you open for a different opinion, a private run, or a safety check.

Straight about the edges

Honest limits

  • Second opinions don't edit files. /also and /compare answer; to have another model do the work, switch to it with /model — the thread carries over.
  • First call to a cold local model waits while the backend loads it — that's the model server, not CloudCode. The next call is quick.
  • The Claude Code bridge is delegation. /claude-cli maps your permission mode onto Claude Code's own layer rather than gating each call itself.
  • It's a terminal tool. Great for engineers and terminal-comfortable vibe engineers — the GUI is where Code Easy comes in.