NewCCode Pro: Claude, GPT, GLM and DeepSeek behind one key

v0.1.8 · macOS · Linux · Windows

The coding agentthat lives inyour terminal.

CCode reads your code, edits it, runs the commands and checks its own work. Any model, you set how much it does alone, and any change can be undone.

$npm i -g @ccode-si/cli
Get started free

Sign up with GitHub for the Free plan: $25 of credits a month.

fig. 01 · Tako, at your service16 × 12 px

// any model, one agent

AnthropicOpenAIDeepSeekZ.aiGoogle GeminiOpenRouterCoreWeaveOllamaLM StudiovLLMllama.cppyour gateway

// see it work

Watch it fix a bug.
Then take the keyboard.

The real interface, in miniature. Once the session ends it's yours: type / for commands and shift+tab to change the permission mode.

  • /rewindputs code and conversation back, without losing either
  • /btwasks a side question that stays out of the chat
  • esc escstops everything at once, helpers included

// side by side

Same model, same prompt.
Different agent.

Every agent ran the same model at default effort, from the same prompt. Every run built something different. Here is what each one made, and what it took.

  • GLM 5.3 Flash
  • CoreWeave
  • effort: default
  • same prompt

>Create an HTML file of a paper boat sailing through a storm.

CCoderec
Agent Arec

1.9×

less context than Agent A

Less to carry means every step costs less.

// time to finish

  • CCode5m 40s
  • Agent A2m 19s

shorter bar: finished sooner

// context at the end

  • CCode17.8k
  • Agent A33.5k

shorter bar: smaller context

Recorded October 2026: GLM 5.3 Flash on CoreWeave at default effort, one prompt per run. Times and context are as each tool reported them in its own terminal.

// what is ccode

A coding agent that
respects your codebase.

The same experience on every model. It works in your terminal, on your files, at your pace.

Read the docs
  • [*]

    Any model, anytime

    Claude, GPT, Gemini, DeepSeek, GLM or a model on your own machine. Switch mid-task and keep going.

  • [*]

    Learns how you work

    Correct it once and it remembers next time. It only keeps what you approve.

  • [*]

    Long sessions stay cheap

    It reuses what the model has already read, so hours of work don't burn through your credits.

  • [*]

    Undo anything

    Rewind your code and the conversation to any point. Nothing you did before is ever lost.

  • [*]

    A team, not a solo act

    Send helpers off to research in parallel while the main task keeps moving.

  • [*]

    Fits your setup

    Add your own skills and tools, connect the services you use, and run your scripts around its work.

  • [*]

    Know what you spend

    Live cost and usage for every session, helpers included. No surprises at the end of the month.

  • [*]

    Safe with any repo

    Clone a stranger's project without worry: it can't change how CCode behaves or where your keys go.

500+

fig. 02 · automated tests behind every release

4

fig. 03 · permission modes, from Auto to Plan

22

fig. 04 · models behind one CCode Pro key

// permission modes

You decide how much it does on its own.

shift + tab cycles them in the app

Reads, edits and runs commands without stopping to ask. Anything you have blocked stays blocked.

~/shop · ccodemode auto

> fix the login redirect loop

●Edit src/auth/redirect.ts

✓Applied · 3 lines changed

●Bash npm test

✓42 passed in 3.1s

  • Edit in the workspace

    yes

  • Shell commands

    yes

  • Outside the workspace

    yes

// long tasks

Big task? It just keeps going.

  • No fresh session needed. Work for hours on one task. When the conversation gets long, CCode tidies it up on its own and carries on.

  • Huge outputs, no meltdown. A giant log or test run won't crowd out your work. It keeps what matters and looks up the rest when it needs to.

  • Never loses the thread. Your latest steps stay exactly as they were, and your full history is always saved.

session memory0% used
auto tidy

reading the codebase…

// pricing

Official model prices.
More of them for your money.

Compare plans

Free

$0/mo

$25 of credits a month

Sign up with GitHub. GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash.

Plan details

Starter

$1/mo

$5 of credits a month

A dollar to try every model.

Plan details

ProPopular

$10/mo

$70 of credits a month

Seven times the usage. For coding every day.

Plan details

// questions

Questions, answered.

Anything else is in the docs.

  • A coding agent for your terminal. It reads your project, edits files, runs commands and checks its work, with permission modes that decide how much it does on its own.

  • Anthropic, OpenAI, Google Gemini, DeepSeek, OpenRouter, CoreWeave, the Z.ai coding plan, or anything that speaks the OpenAI API, including Ollama, LM Studio, vLLM and llama.cpp. Paste a key in ctrl+p → Connect provider.

  • No. Bring your own keys and CCode works with them. A plan gives you one key for Claude, GPT, GLM and DeepSeek models, charged at each model maker's official price.

  • Sign up with GitHub, or connect GitHub to an email account, and you get $25 of credits a month on GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash. Each GitHub account unlocks it for one CCode account. Plan details.

  • macOS, Linux and Windows, on Intel and ARM. One npm command installs it; all you need is Node.js 18 or newer.

  • No. A project you clone can't change how CCode behaves: it can't switch the mode or the model, run its own tools, or reach your keys. CCode also asks before it first opens a folder.

// changelog

Shipped, and still shipping.

npm

  1. v1.0.1

    2026-10-12

    • New/cd <path> moves CCode to another folder without leaving it: that folder's settings, skills, sessions and last open chat. What you had open is saved for the way back.
    • NewEffort levels on CoreWeave for DeepSeek V4.1 Flash, V4 Flash and V4 Pro, GLM 5.3 Flash, GLM 5.2 and Qwen3.8 27B, with "none" to turn reasoning off where the model allows it.
    • NewMore models can see the images they read. On CoreWeave: GLM 5.3 Flash, Kimi K2.7 Code, Kimi K2.6, MiniMax M3, both Qwen models and both Gemma 4 models. Also Claude, GPT-5.1, GPT-5 and Gemini.
    • ImprovedFewer re-reads, fewer tokens. A file changed by a formatter, a test run or your editor no longer has to be read again in full before the next edit.
    • ImprovedEdits that just miss get fixed in one retry. When the text to replace is nearly in the file, the model is shown the line that differs instead of only its own text back.
    • ImprovedTabs are a setting (ctrl+p → Tabs), off by default. Off, CCode shows one session at a time; /new and /session open in its place.
    • ImprovedA working tab's title shimmers, so you can see which sessions are busy.
    • ImprovedCCode credits itself on the commits it writes with a co-author line. Set commitAttribution in ~/.ccode/settings.json to your own line, or to false to turn it off.
    • ImprovedErrors name the file. "Read 3 files · 1 failed: src/missing.ts".
    • ImprovedThe code theme list stays open while you try themes, and code already on screen recolours at once.
    • ImprovedCopying a selection says how much was copied, and one backspace removes a whole collapsed paste.
    • FixedWriting a JSON file could leave [object Object] in it and fail. It is written as formatted JSON now.
    • FixedCCode failed to start in a folder with a very long saved conversation ("Failed to create TextBuffer").
    • FixedA tool call that left out a required parameter failed with a confusing message; the model is now told which one is missing.
    • FixedSearching a provider's name in the model list ("coreweave", "z.ai") shows its models.
    • FixedWhat the model says before writing a file stays above the Write card.
    • FixedContext windows in the model list are right for more models (Gemini, CoreWeave, GLM-5.2, DeepSeek).

  2. v0.1.10

    2026-10-09

    • Newccode mcp: add, list, show and remove MCP servers from the command line.
    • NewA new /mcp view: each server's status and tools, with reconnect and disconnect.
    • NewMonitors. The model can watch a long-running script and react to each line it prints while it keeps working.
    • ImprovedClearer connection errors that say what to check, with a retry countdown.
    • ImprovedSlash commands that open a screen work in the middle of a message; what you typed stays in the chatbox.
    • ImprovedDiffs are easier to read, with deep green and red backgrounds.
    • ImprovedBackground commands write their output to a file, and a finished one is a single line in the chat.
    • ImprovedThe Weights & Biases provider is now named CoreWeave. A saved key keeps working.
    • FixedA command that changed thousands of files could blank the screen. Big changes are now reported by folder.
    • Fixed/undo after a background task's notification removed the wrong part of the chat.

  3. v0.1.9

    2026-10-08

    • New/cloud: run CCode from Telegram. Messages to your own bot run in the tab; answers, permission asks and questions come back with buttons.
    • NewWorks with your editor. Started from the terminal of VS Code, Cursor or Antigravity, CCode follows the file you have open and your selection.
    • New/loop runs a prompt on an interval: /loop 5m check the build.
    • New/goal sets a goal the model is checked against before it stops: /goal all tests pass.
    • New/recap gives a one-line recap of the session.
    • NewTransparent theme: your terminal's own background shows through.
    • NewWrites and edits show as soon as the model starts them.
    • ImprovedConnected providers are refreshed at every start: new models appear, dropped ones go.
    • ImprovedThe / menu opens mid-text, finds commands by their description, and shows a hint for commands that need more.
    • ImprovedWeb searches and fetches read like calls; click the line to see the results.
    • FixedLong conversations with Claude cost far more than they should. The prompt cache now keeps hitting.
    • FixedBig command output no longer floods the context. Output over 30,000 characters goes to a file the model can search.
    • FixedEffort on current Claude models: /effort now works where it showed "Not configurable".
    • Fixed/rewind and /fork no longer reset the "$ spent" and token totals.
    • FixedThinking times and the "done" line under each answer survive reopening a session.

Your next commit starts in the terminal.

Install it, run ccode in a project, and describe the task.

$npm i -g @ccode-si/cli
Create an account