What Is Agent Reach? Giving an Agent Internet Access, Installation, Cookies and Data
You install Agent Reach, and your AI agent becomes able to read sources such as web pages, YouTube, GitHub, RSS, X/Twitter, Reddit, Instagram, Facebook and LinkedIn. Agent Reach doesn't do the reading itself: it installs the tools needed (gh, yt-dlp, OpenCLI), checks their health (doctor) and points the agent to the right tool. Note: the default `agent-reach install --env=auto` only checks and installs nothing on your system; installation and writing the skill happen only with the `--system` flag. Platforms that require sign-in use cookies or an open Chrome session, and the account may be restricted; the documentation recommends a secondary account. Content that's read goes to third parties (Jina Reader, Exa, Groq). MIT licensed.
- What it does
- Sets up the agent to read the web, YouTube, GitHub and social networks
- Type
- CLI + skill (Python package)
- Where
- Claude Code, OpenClaw, OpenCode and agents that can run commands
- Requirements
- Python 3.10+, pipx; extra tools depending on the channel
- Cookies
- Platforms that require sign-in use cookies or a Chrome session; the account may be restricted
- Turkish
- Not stated in the official documentation
- Cost
- Free (MIT); possible proxy cost on restricted networks
What Is Agent Reach Used For?
It installs the tools an agent needs to reach the internet for tasks like "get this video's subtitles," "search this GitHub repository" or "read this page," checks whether they work (agent-reach doctor), and leaves a skill file that tells the agent which tool to call. It doesn't do the reading itself; the agent calls the upstream tools (gh, yt-dlp, OpenCLI, twitter-cli and similar) directly.
| Platform | Account or cookies | Backend |
|---|---|---|
| Web, RSS, V2EX | None | Jina Reader, feedparser, public API |
| YouTube, GitHub | None for public content | yt-dlp, gh CLI |
| Web search | Keyless endpoint | Exa (via mcporter) |
| Twitter/X | auth_token and ct0 cookies | twitter-cli, OpenCLI as fallback |
| Reddit, Facebook, Instagram | Session required; signed-in Chrome or cookies | OpenCLI (rdt-cli as fallback) |
| Manual sign-in in the browser | mcporter and mcp-server-linkedin, Jina Reader | |
| Xiaohongshu, Bilibili, Xueqiu, podcasts | Cookies or a Groq key, depending on the platform | OpenCLI, bili-cli, cookie-based API, Groq |
The README table counts 15 platforms, while SKILL.md says 16 (a job listings platform was added). Who it's for: people who want to do internet research with their agent. This means the installation will run external tools on your computer; read the "What does it access?" section below.
How Do You Install Agent Reach?
Requirements: Python 3.10 or later, and pipx (recommended) or pip. The commands are taken verbatim from the README and the installation guide. The "agent-reach" package on PyPI isn't this project; you install it from the GitHub zip URL.
› pipx install https://github.com/Panniantong/agent-reach/archive/main.zip› agent-reach install --env=autoThe installation guide says this command is a check only by default ("Read-only check (default)"): it looks at the gh, Node.js, mcporter, Exa and yt-dlp setup and doesn't change your system.
› agent-reach install --env=auto --dry-run› agent-reach install --env=auto --system--system installs and configures the missing pieces and writes the skill file to the agent's folder. The installation guide says to use it only after the user has explicitly approved. There's an option like --channels=opencli,xiaohongshu for specific channels only.
If gh CLI and Node.js are missing, --system installs them itself only on Linux (apt-get) and macOS (Homebrew); on Windows you need to install them manually. There are open issues for Windows (#566, #631, #590, #586).
› npx skills add Panniantong/Agent-Reach@agent-reachOnce the skill is installed, the agent checks whether the agent-reach command is installed and installs it if not (README). The README also suggests giving the agent the URL of the installation guide and letting it do the install; with this method, look at what the agent will run beforehand.
› py -3 -m venv $env:USERPROFILE\.agent-reach-venvThen activate the virtual environment and run python -m pip install https://github.com/Panniantong/agent-reach/archive/main.zip and agent-reach install --env=auto. If python3 opens the Microsoft Store on Windows, use py -3.
To verify the installation, agent-reach doctor prints the channel status. On platforms that require sign-in, doctor deliberately doesn't run a live check (to avoid reading cookies), so you may see "active_backend: null"; do the real test with your agent and look at the result.
How Do You Use Agent Reach?
Once the skill is installed, you write your request normally; the agent checks the skill and picks the right tool.
› agent-reach check-update!The documentation says automated requests with cookies on platforms that require sign-in can get the account restricted, and it recommends using a secondary account and waiting 2-3 seconds between requests.
Which Agents Does Agent Reach Work With?
| Claude CodeYes | The skill is written to the ~/.claude/skills folder if it exists (per the source code). The README says "any agent that can run commands." |
|---|---|
| OpenClawYes | ~/.openclaw/skills; exec permission must be enabled (README). |
| OpenCode and the generic ~/.agents/skillsYes | Listed in the code; if none of them exist, ~/.agents/skills is created. |
| Windows (end to end)Unclear | The documentation doesn't say it has been tested; there are open Windows issues. |
What Does Agent Reach Access? Does It Read My Cookies?
The default installation doesn't read browser credentials (in the code: "install never reads browser credentials"). Cookies are only taken when you ask: a manual export with Cookie-Editor, or a single platform's cookies from your local browser with agent-reach configure --from-browser chrome --platform <platform>. The documentation says the third-party rdt-cli's rdt login command extracts the browser cookies on its own. The OpenCLI route uses your open session through an extension you install in Chrome, not a cookie file.
~/.agent-reach/config.yaml (cookies, tokens, proxy; file permission 0600 except on Windows). With --system it writes the skill folder, an Exa entry in mcporter, global npm packages (mcporter, OpenCLI), pipx tools and one line in the yt-dlp config.
SKILL.md doesn't count write actions such as posting, commenting and liking as part of this skill's job, but references/dev.md teaches gh issue create, gh pr create, gh repo create and fork examples. The documentation contradicts itself; if gh is signed in, the agent can run these.
No telemetry code appears in the source; check-update only sends requests to api.github.com. But content goes to third parties: the web address being read goes to Jina Reader, web searches to Exa, podcast audio to Groq. The README says cookies stay local and aren't uploaded.
!Untrusted page content can look like instructions to the agent. The uninstall command deletes settings, cookies and skill files. The repository has no published security advisory; an open issue (#802) pastes a "critical" report from an automated scanner called NVIDIA SkillSpector, but there's no maintainer response and it isn't a verified finding. This doesn't mean "safe"; since the installation runs many external commands, check what it will do with --dry-run first.
Is Agent Reach Free?
Yes. It's MIT licensed; the README says "completely free" and mentions a proxy of about one dollar a month on restricted networks as the only possible cost. Podcast audio requires a Groq key that the documentation calls free; check the current terms on Groq's own page. The sponsor links in the README are paid third-party products and aren't required. The cost of the model your agent uses is separate.
- MIT: free to use and modify
- The platforms' own terms of use apply: automated access with a signed-in session can get your account restricted
Source: README and installation guide. This is not legal advice.
Common Problems and Fixes
| What's happening | What to do |
|---|---|
| pip "externally-managed-environment" (PEP 668) error | Use pipx, or create a virtual environment with python3 -m venv ~/.agent-reach-venv. |
On Windows, python3 opens the Microsoft Store | Use py -3 or the real path to python.exe. |
| Reported by users On Windows, some channels, LinkedIn and transcribe.sh don't work | Open issues #631, #590, #586, #566; the documentation has no definitive fix. |
| doctor says "active_backend: null" on platforms that require sign-in | By design: doctor doesn't run a live check, to avoid reading cookies (this is criticized in issue #566). Have it actually do the task and verify the result. |
| Reported by users I updated, but the version didn't change | Open issue #803: main is about 99 commits ahead of the tag, but the version string is still 1.5.0; pip sees the same version and doesn't update. |
| Your account was restricted | The documentation's own warning: cookie-based automation can get an account restricted. Use a secondary account and wait between requests. |
How Do You Uninstall Agent Reach?
- Preview:
agent-reach uninstall --dry-runshows what will be deleted. - Uninstall:
agent-reach uninstalldeletes the configuration, cookies, skill folders and the Exa and Xiaohongshu entries in mcporter.--keep-configkeeps the configuration. - Remove the package:
pipx uninstall agent-reach(pipx's standard command; the Agent Reach documentation doesn't list it separately). - Upstream tools: Remove separately installed tools like twitter-cli and rdt-cli yourself; uninstall reminds you of this but doesn't do it.
To make sure the cookies become invalid on the platform side, it makes sense to sign out of that platform; the documentation doesn't say this, it's general security advice.
Frequently asked questions
What is Agent Reach?
An MIT-licensed capability layer that lets an agent read internet sources (web, YouTube, GitHub, RSS, X, Reddit, Instagram, Facebook, LinkedIn and more): it handles installation, health checks and routing; the upstream tools do the reading.
Does the installation read my browser cookies?
The default installation doesn't. Cookies are only taken when you ask: a manual export, or configure --from-browser for a single platform. The third-party rdt-cli's rdt login command, however, extracts the browser cookies on its own. The OpenCLI route uses your open Chrome session.
Can my agent post or comment on my behalf?
The skill doesn't count this as part of its job, but on the GitHub side it teaches examples of creating issues, PRs and repositories with gh; if gh is signed in, the agent can run them. Keep Claude Code's permission prompts on.
Can I install the agent-reach package from PyPI?
No. A commit in the repository says the package with the same name on PyPI is an unrelated project; you install it from the GitHub zip URL.
Is it safe?
This isn't a security endorsement. The repository has no published security advisory. An open issue (#802) pastes a critical report from an automated scanner; there's no maintainer response and it isn't a verified finding. Check what it will install with --dry-run first.
Related guides
Sources
Not clear in the official sources: Whether Agent Reach works end to end on Windows · the privacy terms of the third-party tools (OpenCLI, rdt-cli, twitter-cli, mcporter) and of Exa, Jina and Groq · whether the Groq key is free (stated only in the Agent Reach documentation) · the accuracy of the scanner findings in #802 · whether Claude Code recognizes the ~/.claude/skills path (only in the code). Each one is covered in its section.