raw-cli
A harness for your own agent: tool loop, sessions, approvals, REPL and dashboard, with tools and skills kept as editable files.
- 108 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The `bash`, `write_file` and `edit_file` tools run without a prompt with your account permissions, and `cwd` does not isolate files.
- The dashboard with the `--remote` flag exposes the local agent interface on a phone through a Cloudflare tunnel.
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/raw-cliDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The `bash`, `write_file` and `edit_file` tools run without a prompt with your account permissions, and `cwd` does not isolate files.
- The dashboard with the `--remote` flag exposes the local agent interface on a phone through a Cloudflare tunnel.
This is third-party code. Review the repository files before installing.
What it does
raw is a CLI with a ready harness for building your own agent: a tool-call loop, context compaction, saved sessions, approvals, a terminal REPL, a local browser dashboard and the ACP protocol for editors. Everything the agent does lives in files you can change: tools, Markdown skills, hooks and access rules. Fourteen plugin tools ship in the box, including `read_file`, `write_file`, `edit_file`, `bash`, `apply_patch`, `todo` and `process`. Models connect through config over the OpenAI, Anthropic and Google APIs, including local ones through Ollama. Selected MCP servers add tools, and a finished agent with its tools and skills is packed into a `.rawpkg`.
Who it is for. Developers building an internal, coding or ops agent who want to change parts of a ready harness without writing the loop from scratch.
Good fit when
- An internal agent needs its own tools and access rules.
- A model runs locally through Ollama and gets a minimal tool set.
- An agent with its tools and skills is handed to colleagues as a `.rawpkg` file.
- An editor connects the agent over the ACP protocol.
Not a fit when
- A sandbox is needed: `cwd` only resolves relative paths and isolates nothing.
- A ready assistant is wanted without configuring models and a config file.
- The machine has Node.js older than 22.13.
Example request
Run `raw --agent raw "Explain the tests in this repository"` and show which files the agent read.Limitations
Node.js 22.13 or newer and Bash for the `bash` tool are required; on Windows without Git Bash the `powershell` tool is selected. Tool calls run automatically with your OS account permissions, and `cwd` is not a sandbox. Without an `ask` or `deny` rule in `tools.rules` a command runs without a prompt. Context compaction and caching depend on the model and API, and image support is enabled per model. The cost of paid API requests falls on your key.
How to disable. Remove the package with `npm uninstall -g @tlelabs/raw`. The config `~/.config/raw/` and the session directory `~/.local/state/raw/` sit in the home directory by default and are deleted by hand.
FAQ
How do I install and run raw?
`npm install -g @tlelabs/raw`, then `raw config init` creates `~/.config/raw/config.json` with the `raw` agent. For a one-off run without installing, use `npx @tlelabs/raw --help`.
Which models can I connect?
Any model with one of four API methods: `openai-chat-completions`, `openai-responses`, `anthropic-messages` or `google-generate-content`. The config holds `model_id`, `base_url` and `api_key_env`.
How do I limit what the agent does?
Through the agent `tools.rules`. The `allow`, `ask` and `deny` rules match a tool or a pattern, for example an `rm` command in `bash`. The `-y` flag does not bypass an explicit `ask` rule.
How do I hand an agent to a colleague?
Export the agent with `raw package export`, pack it into a `.rawpkg` with `raw package pack` and install it on another machine with `raw package install`. The recipient model is bound with `raw agent add`.
What does the dashboard show?
`raw dashboard` opens a local browser app with sessions, tool calls, context usage, approval requests and settings. The `--remote` flag also opens it on a phone through a Cloudflare tunnel.
Discussion
No questions or reviews yet.
Related
CLI toolsA headless coding agent in one Go binary: reads and edits files, runs commands and never trims context silently
A scaffold of personas, rules and skills in plain Markdown: Maestro hands work to different models and spares your main plan
An open coding agent for the terminal, browser and desktop that works with models from OpenAI and Claude to Ollama
One runtime for Claude Code, Codex, OpenCode and more: a run counts as done only when checks the agent did not write pass
rsc-harness is a CLI for configuring agent skills and project rules through a plan the user accepts before files change.
A teaching terminal coding agent in Python, built as a "nanoGPT for agents": read it, take it apart and fork it
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| raw-cli (this card) | no | 108 | |
| MiniCode | no | 1,162 | a compact terminal coding agent with a TUI; raw-cli adds sessions, a dashboard, approvals and agent packaging |
| MetaHarness | no | 738 | builds an agent package from a repository; raw-cli provides a harness where tools and rules are edited as files |
README in short
Documentation lives at raw.tlelabs.com, and the repository has guides for configuration, tools, skills, MCP, hooks, ACP and the dashboard. Tool examples and a ready `project-helper` agent ship in the package.