Carbon Code
A DeepSeek-powered coding CLI that decomposes tasks, proposes code changes, and validates them with tests.
- Source repo
- Yapie0/carboncode
- Stars
- ★ 199
- Last updated
- 3mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 56/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI APICodex (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A DeepSeek API user who wants an interactive terminal agent to inspect a local repository, propose diffs, and help implement changes.
- Not a fit if
- Developers on Node.js versions older than 22
- Users unwilling to obtain a DeepSeek API key or pay model usage costs
- Website
- yapie0.github.io/carboncode ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-08-10
- Source review
- 56/100 · Major gaps
What does this agent do, and when should you use it?
Carbon Code is a TypeScript terminal coding agent whose default model presets target DeepSeek V4. Launched inside a project, it reads and searches the codebase, proposes changes as diffs, requests approval before shell execution, and runs tests when needed. Its interfaces include interactive coding, tool-free chat, one-shot non-interactive tasks, and project-guidance generation. The CLI installs the carboncode and ccode commands, stores user configuration in ~/.carboncode/config.json, and recognizes repository guidance in AGENTS.md or CARBON.md. A desktop client also supports standard OpenAI-compatible providers, including model discovery through /models and adaptation between the Responses and Chat Completions APIs. It is aimed at developers who want an agent operating against a local repository while retaining an approval checkpoint for commands.
A typical session starts with carboncode or carboncode code [dir]. The agent reads and searches project files, decomposes the task, proposes edits, and presents the resulting changes as a diff. It asks for confirmation before running shell commands, can execute tests to validate the work, and leaves a concise session record. carboncode run "task" performs a single non-interactive task, while carboncode chat removes filesystem and shell tools. carboncode init [dir] reads manifests, directories, and tool configuration to generate CARBON.md without calling a model. The documented command set also includes carboncode doctor, carboncode update, MCP testing, and collaboration with agents such as Claude and Codex. Model presets map flash to deepseek-v4-flash, pro to deepseek-v4-pro, and auto to Flash with a one-time escalation to Pro on a difficult turn.
- A DeepSeek API user who wants an interactive terminal agent to inspect a local repository, propose diffs, and help implement changes.
- A team that needs a one-shot coding task in a script or automation flow through
carboncode run "task". - A maintainer onboarding to an existing project who wants to preview and generate a
CARBON.mdguide withcarboncode init --dry-run. - A developer who wants agent-run tests and shell commands but requires an explicit confirmation step before command execution.
- A team evaluating MCP tests or workflows that coordinate Carbon Code with Claude or Codex.
- A user who only wants a conversation and does not want the session to receive filesystem or shell tools, using
carboncode chat.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npm install -g @carboncode/clinpx @carboncode/cli
- Environment variables
DEEPSEEK_API_KEY- Config files
~/.carboncode/config.json- Runtime
- Node.js 22
- Install notes cover
- Windows
How do you install or deploy this agent?
Install Node.js 22 or newer and obtain a DeepSeek API key.
Install globally and start it in a project:
npm install -g @carboncode/cli
cd path/to/project
export DEEPSEEK_API_KEY=sk-...
carboncodeThe installed shorthand is ccode. To avoid a global installation, run npx @carboncode/cli. If PowerShell blocks npm.ps1, use npm.cmd install -g @carboncode/cli. The first-run wizard can also configure the DeepSeek API key, and persistent user settings live at ~/.carboncode/config.json.
How do you use this agent?
Run carboncode in the current project or carboncode code [dir] for a specified directory. Review the proposed plan and diff, then approve or reject shell-command requests and allow tests when appropriate. Use carboncode run "task" for one non-interactive job, carboncode chat for a session without filesystem or shell tools, and carboncode doctor to check the local environment. For project guidance, run carboncode init --dry-run before carboncode init; existing guidance files are preserved by default, while carboncode init --force --yes explicitly permits replacement. Run carboncode update to check for and install the latest CLI package. In the desktop client, an OpenAI-compatible provider can be added under Settings -> Models -> Add model provider by entering its Base URL and API key.
What are this agent's strengths and limitations?
- Covers a concrete local-development loop: code search, task decomposition, diff review, command approval, and test validation.
- Offers interactive coding, tool-free chat, and one-shot
carboncode runmodes for both manual and scripted workflows. carboncode initdoes not invoke a model and includes dry-run and no-overwrite-by-default safeguards.- The desktop client supports standard OpenAI-compatible providers, discovers
/models, and adapts between Responses and Chat Completions APIs. - MCP testing and collaboration with Claude and Codex are explicitly identified as supported capabilities.
- The CLI requires Node.js 22 or newer, creating an upgrade requirement for older environments.
- The default provider path requires a DeepSeek API key; the claimed 90%+ token-cost saving is not accompanied by benchmarks, methodology, or billing evidence in the supplied material.
- Desktop installation and deployment instructions are absent from the supplied material, even though provider configuration is described.
- The material does not provide setup steps, support boundaries, or validation examples for MCP testing and Claude/Codex collaboration.
- Redacted error and fatal diagnostics are uploaded by default, so organizations with strict telemetry policies must disable the feature and account for its offline pending-event directory.
How does this agent compare with similar options?
The project positions Carbon Code as similar to Claude Code and says its capability approaches Claude Sonnet 4.6, while emphasizing DeepSeek as the default provider and claiming more than 90% token-cost savings. The supplied material contains no matched-task evaluation, quality benchmark, latency results, or cost methodology, so those capability and cost comparisons remain project claims rather than independently supported measurements. Carbon Code also documents standard OpenAI-compatible provider support in its desktop client.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Carbon Code This agent | 56 · Major gaps | CLIFree + model costs | ★ 199 | 3mo ago | TypeScript | OpenAI API |
| Deep Copilot | 60 · Some gaps | IDE extensionFree + model costs | ★ 79 | 5mo ago | JavaScript | — |
| Chat On Steroids | 76 · Good | Desktop appFree + model costs | ★ 5.2k | today | TypeScript | ChatGPT |
| Kode CLI | 60 · Blocked | CLIFree + model costs | ★ 5.2k | 14d ago | TypeScript | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The docs frame the tool for personal developer workflows and describe shell approval, edit review, and sensitive-path configuration. Diagnostics documentation says redacted error metadata and stacks are collected by default, lists excluded content, and provides an opt-out. Deductions: the CLI sample exposes clues to broader modes such as --yolo and collaboration, while the supplied material does not detail every tool permission boundary, network data flow, credential-store protection, or recovery path; dependencies are listed by version range without security audit evidence.
The CLI documentation lists commands and a Node 22 requirement; the test samples cover behavior such as continuing after MCP initialization failure, and CI configuration specifies build, static checks, and tests. Deductions: samples do not establish whole-repository results, external npm, model API, and MCP availability depend on outside services, and no execution was performed.
The README identifies personal developers as the audience and documents interactive coding, filesystem-free chat, one-shot tasks, project initialization, and desktop support for compatible providers; command documentation and the Node requirement help users assess fit. Deductions: permission boundaries and limitations are only partly described, and the supplied sample does not establish complete platform or scenario coverage.
The README, command table, configuration, installation notes, and license information are clearly organized. Package metadata gives a stable name, version, repository, and issue route, while MIT licensing and third-party notice locations are explicit. Deductions: no changelog is included in the packet; limitation and security-boundary documentation is incomplete; maintenance responsibility can be inferred from the issue route and SECURITY.md but no team or response commitment is stated.
The evidence describes reading and searching project code, task breakdown, editing, and running commands after approval; the library exports also show a broad agent, tool, and MCP surface suited to coding workflows. Deduction: the claimed savings above 90% and capability near Sonnet 4.6 are unsupported descriptions here, without measurement method, comparison, or cost data.
README installation, command, and configuration details correspond in part with package metadata and CLI definitions, and selected MCP behavior has test examples. Deductions: the claimed cost savings and model capability lack quantitative evidence; cross-source corroboration of core claims is limited, and static materials cannot establish that tests were actually executed.
- Redacted error metadata and stack traces are collected by default; disable this in Desktop settings or through configuration/environment if undesired.
- The claims of over 90% token savings and capability near Claude Sonnet 4.6 are not supported by measurements or comparisons in the supplied material.
- This static review did not execute code or tests; review shell, edit, MCP, and --yolo permission settings before use.
FAQ
What credentials are required?
DEEPSEEK_API_KEY. An OpenAI-compatible provider in the desktop client requires that provider's Base URL and API key.Can it execute local commands without asking?
carboncode chat.Does diagnostics upload source code or conversations?
CARBONCODE_DIAGNOSTICS=off or {"diagnostics":{"enabled":false}}.What happens to diagnostics when the network is unavailable?
~/.carboncode/diagnostics/pending/ with count and size limits. The supplied material does not state the numerical limits.Will project initialization overwrite existing guidance or call a model?
carboncode init only reads repository manifests, directories, and tool configuration and does not call a model. Existing guidance files are not overwritten by default; use --dry-run first and pass --force only when replacement is intended.Is Carbon Code free?
Is Carbon Code open source?
Is Carbon Code safe to use?
Is Carbon Code still maintained?
Who makes Carbon Code?
What is Carbon Code's official website?
Thanks — we use this to improve the page.