Waveloom CLI
Terminal coding agent in Go for DeepSeek, Kimi, GLM and OpenAI, built around prefix caching and confirmed writes and commands
- 138 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- Executes shell commands and writes files, with each action going through confirmation, but allow rules can remove the prompt
- Stores API keys in ~/.waveloom/ and reaches the network through web_fetch and web_search
Install
- Works in
Manual install
curl -fsSL https://raw.githubusercontent.com/Menfre01/waveloom/main/install.sh | shmacOS and Linux, installs to ~/.local/bin with no sudo. On Windows there is install.ps1 and Git for Windows is required.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Executes shell commands and writes files, with each action going through confirmation, but allow rules can remove the prompt
- Stores API keys in ~/.waveloom/ and reaches the network through web_fetch and web_search
This is third-party code. Review the repository files before installing.
What it does
Waveloom runs a coding agent right in the terminal. The architecture keeps the longest common prompt prefix hot in the DeepSeek cache across turns, so repeated input costs noticeably less. The model picks the mode itself: pro for deep reasoning and flash for routine work. A fixed system prompt, append-only history and four-tier watermark compaction keep the prefix unchanged. It ships 14 tools: read, write, edit, bash, web_fetch, web_search, skill and sub-agent calls, plan mode and todo lists. Skills plug in like in Claude Code: SKILL.md files with nine frontmatter fields are read from .claude/skills/ and .claude/plugins/. There is a full MCP client, PreToolUse, PostToolUse, Notification and Stop hooks, Evaluate, Explore and Verification sub-agents, checkpoints with file rewind and session continuation through waveloom --continue. File writes and command execution require your confirmation. Permission decisions follow allow, deny or ask rules, and tool output passes five checks including a scan for injected instructions. An optional sandbox runs on bubblewrap on Linux and Seatbelt on macOS.
Who it is for. Developers who run long sessions on DeepSeek and count token costs but want familiar .claude/skills and a full terminal interface.
Good fit when
- A long DeepSeek session where the bill for cache-missing input keeps growing
- The project already holds Claude Code skills and you want to run them on another model without rewriting
- You switch between DeepSeek, Kimi, GLM and OpenAI providers with /provider
- Before changing code you need plan mode where only the plan file can be written
- A screenshot or chart has to be analyzed by a multimodal model: @shot.png
Not a fit when
- You use Anthropic and need coordinator mode: the author itself recommends Claude Code
- You need a desktop GUI and QQ bot integration: Waveloom is terminal-only
- Windows without Git Bash: bash.exe is required and cmd and PowerShell are not supported
- The model has to go through a third-party relay: the key goes straight to the provider
Example request
Switch to plan mode, explore the authorization module and propose a refactoring plan, writing nothing except the plan fileLimitations
- An API key for DeepSeek, Kimi, GLM or OpenAI is required: it is stored in ~/.waveloom/ and goes straight to the provider.
- On Windows Git for Windows is required, only Git Bash and WSL2 work and cmd and PowerShell are not supported.
- Web search through the deepseek provider is billed separately: 5 CNY per 1000 successful calls.
- Automatic post-edit LSP checks are declared for Go, Rust, TypeScript, JavaScript and C/C++, while other languages use build commands such as go build.
How to disable. Delete the binary from ~/.local/bin (brew uninstall waveloom if installed through Homebrew) and the ~/.waveloom/ directory with settings and keys.
FAQ
How do I install Waveloom?
On macOS and Linux run brew install menfre01/tap/waveloom or the script curl -fsSL https://raw.githubusercontent.com/Menfre01/waveloom/main/install.sh | sh. Then waveloom setup and waveloom.
Does Waveloom pick up Claude Code skills?
Yes, SKILL.md is read from .claude/skills/ and installed plugins from .claude/plugins/. Nine frontmatter fields are supported, including $ARGUMENTS and paths.
How do I switch models or providers?
/model or waveloom --model deepseek-flash changes the model, /provider or /provider kimi changes the provider. Profiles live in settings.json under llm.profiles.
Is my API key safe?
The key is stored locally in ~/.waveloom/ and goes directly to DeepSeek, Kimi, GLM or OpenAI with no third-party relay.
How do I return to a session days later?
Run waveloom --continue: the agent remembers the prior context. Checkpoints let you rewind files to any message.
Discussion
No questions or reviews yet.
Related
CLI toolsA terminal coding agent in a single Go binary for DeepSeek and other OpenAI-compatible models, with a code graph and local OCR
A local coding agent with a choice of provider: DeepSeek, Kimi, OpenAI or Anthropic. Terminal and browser interfaces.
A Python terminal coding agent in the manner of Claude Code: reads the repository, edits files, runs commands, with memory, sub-agents and MCP
A terminal coding agent tuned for the deepseek-v4 model: deep thinking, reasoning effort control, Agent Skills and MCP
A terminal coding agent in Go: GLM, DeepSeek and local Ollama models, 59 built-in tools and MCP
Lightweight Python terminal coding agent: ReAct loop, file-based context, guarded bash and skills, works with DeepSeek and OpenAI
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Waveloom (this card) | no | 138 | |
| Claude Code Harness | no | 3,207 | a harness around Claude Code; Waveloom is an agent with its own loop and DeepSeek cache design |
| DeepSeek Harness | yes | 280,265 | another DeepSeek harness; this one adds four-tier compaction and Claude Code skill compatibility |
README in short
The page presents Waveloom as a terminal agent built around DeepSeek cache economics and shows a demo. Then come install for macOS, Linux and Windows, configuration, a comparison table with Claude Code and Reasonix, highlights, answers to questions and a list of documents on settings, cache, MCP and IDE connection.