Agent Reach is an open-source CLI and skill that gives AI agents a common way to search and read the web, YouTube, GitHub and other platforms. Its official 1.6 setup uses uv, followed by agent-reach setup, to prepare default channels and agent instructions. Start with the agent's existing public search tools; use Reach when the task needs missing text, captions or a declared platform action.Sources: installation guide, skill routing rules
What does it add to an agent?
Reach puts upstream tools, fallback routes and structured errors behind agent-reach <channel> <action>, with read choosing a route from a URL. It is a third-party search and reading tool, rather than an official authorization gateway for these platforms. Its current skill permits reading only: no publishing, commenting, liking or direct messages. The skill tells agents to use adequate public results first; a platform name or link alone should not trigger installation or login.Sources: README, skill
Editorial judgment 1: define the evidence before choosing the platform. Finding five relevant videos can use titles, links and descriptions. Comparing what their speakers say requires caption or transcript text. The official capability guide distinguishes a caption-track list from captions themselves. Defining the output first can avoid unnecessary login and maintenance work.Basis: capability guide
Default, optional and experimental channels
This is the v1.6.1 classification. Default means configured by setup; it does not mean every action needs no login or key, or works on every network. A login entry is not a prerequisite for all public content. Channels that have not been added can try a public-web route, with a potentially narrower result.Sources: channel matrix, routing rules
| Class | Channels | Scope and prerequisites |
|---|---|---|
| Default | web, github, v2ex | Web search/read, public GitHub repositories and issues/PRs, V2EX topics and replies; public access does not establish private-repository permission |
| Default | youtube, bilibili | Search, video information and captions; Bilibili captions need login, while YouTube speech transcription needs a Groq or OpenAI key |
| Optional | xiaohongshu, reddit, boss | On-platform notes, posts/comments or jobs; the matrix lists login requirements, but assess whether a public fallback already satisfies the task |
| Optional | linkedin, xueqiu | Public LinkedIn jobs can be read first; on-platform search needs login. Xueqiu quotes need no login, while community hot posts/stocks do |
| Optional | xiaoyuzhou | Podcast transcription with a Groq key |
| Experimental | twitter, facebook, instagram | Declared search/profile actions with login; upstream warns that platform defenses can break these routes |
Support also has content limits. YouTube and Bilibili have no unified comments action. Xiaohongshu read returns a text description, not evidence of understanding its images or video. Reddit does not promise a complete comment tree. Judge the returned text, source and scope; a search snippet is not full content.Source: capability guide
Install and make a first query
The package requires Python 3.10 or later; the 1.6 series uses uv. Read the official guide and confirm terminal execution and uv availability before setup. The PyPI package named agent-reach is not this project. Do not substitute pip install agent-reach for its official GitHub installation route.Sources: package metadata, installation guide
The core official commands are below. The latest URL follows future releases; at this page's verification date, the latest release was v1.6.1, published November 1, 2026.Sources: installation guide, v1.6.1
uv tool install https://github.com/Panniantong/Agent-Reach/releases/latest/download/agent-reach.zip
agent-reach setup
agent-reach channels
agent-reach doctor --text
setup enables default channels, prepares tools and installs the skill. It is not a dry run that only lists changes. Adding an optional channel can also prepare Node, OpenCLI, a browser engine or an extension; the Windows Chrome fallback may require the user to confirm an extension. Read the generated installation report and outstanding actions instead of treating installation as proof that every action works.Sources: installation guide, 1.6.0 changes, session guide
These examples follow the official command format; inspect channels for available actions and parameters. The returned ok, content items and error.hint separate the outcome from the next step.Sources: README, skill
agent-reach bilibili search "大模型"
agent-reach read https://www.v2ex.com/t/123
For 1.5 or earlier, the official migration guide uses a forced uv installation, a version-output check, then setup. Do not reuse the old page's install --system, doctor --live or standalone RSS-channel instructions. The 1.6 rebuild changed the entry points and removed the RSS channel.Sources: migration guide, 1.6.0 removals
What do doctor and check establish?
agent-reach doctor --text displays offline status, tool versions and saved check records. It does not reconnect to establish current platform access. agent-reach check <channel> --text performs network checks. A search-only check does not establish that reading, captions or every other action also works.Sources: CLI definitions, status definitions
unverified: configured, but no live check in the last 24 hours.search_ready: a search returned nonempty results in the last 24 hours; other actions need separate checks.ready: passed the recorded checks in the last 24 hours; inspect which actions were covered.regressed: true: a channel flag showing that an action passed the previous check but failed the latest one; it is not another status name.
Editorial judgment 2: separate installation from evidence completeness. A market brief needs source URLs and dates; a video comparison needs caption or transcript text; a social-listening summary should state the comment scope. An installed tool, saved cookie or search_ready status does not justify claiming the full content was read.Basis: capability guide, status definitions
Before using signed-in channels
How are login sessions handled? Browser channels share Reach's dedicated browser profile. On first use of a site, if the user's everyday Chrome is connected to the OpenCLI extension, the session guide says the program copies that site's cookies to the dedicated browser. It does not copy the whole browser, other sites or the password database, and leaves the source browser unchanged. Without a connected source, automatic reuse is not guaranteed; saved cookies do not prove a session remains valid.Sources: session guide, session-reuse implementation
Confirm the account before using it. The copied login may be the everyday main account. Upstream recommends a dedicated secondary account because platforms may restrict automated access. A dedicated browser does not remove account risk. Treat cookies as login credentials and include local permissions, backups and shared logs in their handling. When login is actually needed, the user completes it on the official page; passwords, cookies and keys should not be pasted into the agent chat.Sources: README safety guidance, session guide Local-data handling is our editorial recommendation.
For RATE_LIMITED, wait for the specified interval. For PLATFORM_BLOCKED, stop and have the user handle verification on the official page; do not rotate accounts, backends or browsers to keep requesting. NEED_LOGIN is not a reason to force a login whenever only a snippet is available: first establish whether the remaining task needs signed-in content.Source: skill error handling
Cost, updates and known problems
The code uses the MIT license. That does not establish the price of models, transcription, proxies or platform services. YouTube speech transcription needs a Groq or OpenAI key; Xiaoyuzhou needs Groq. Charges depend on the selected service and usage. Access also does not settle platform terms, privacy or permission to reuse content.Sources: MIT license, channel matrix Cost and reuse checks are editorial guidance.
The v1.6.1 release records fixes for empty Xiaohongshu content despite a success response, a Bilibili login that was not saved, Windows Xiaohongshu URLs containing &, and browser-port conflicts. These are upstream change records, not a guarantee for every account, network or item. If another tool owns port 9222, the official example moves Reach to agent-reach configure browser-port 9333, instead of closing the other tool. Failed login copying should produce an error; copying zero cookies should not count as success.Sources: v1.6.1, browser-port guide, fix PR #879
Next steps
- Before asking Codex to install tools, review folders, network access and approvals.
- For your own mail and cloud documents, compare Google Workspace CLI's OAuth and service scopes.
- For AI coding endpoints and providers, see CC Switch.
Sources and verification date
Verified November 5, 2026. Release baseline: v1.6.1, published November 1, 2026, with v1.6.0 for the rebuild history. This guide synthesizes official documentation, source code and release records; the diagram and two work judgments are original editorial work.