Agent Reach - Internet Access for Agents
Keep and recommend as the lightweight multi-platform research router. It wires agents to channel-specific tools and doctors rather than hiding every source behind one opaque runtime.
Routing Summary
Use Agent Reach when a research task crosses live public platforms such as X, Reddit, YouTube, GitHub, RSS, LinkedIn, Bilibili, or Xiaohongshu. Prefer a zero-auth public route first, configure only the platform required, and move to Browser Testing Skills only when interaction is unavoidable. Use Firecrawl for structured extraction, multi-page crawling, monitoring, or JavaScript-heavy sites after direct/Jina/Defuddle - Web Content Extraction routes are insufficient.
Do not bulk-import browser cookies, use a primary account when a dedicated research account works, or assume “zero API fees” means zero account, terms, rate-limit, or privacy risk.
Current Source Snapshot
| Surface | Reviewed state |
|---|---|
| Repository | Panniantong/Agent-Reach at 1221ecd0c3e0502ee37406f03543bedf7503f2c7 |
| Version | v1.5.0, Python >=3.10, MIT |
| Reach | 70,317 stars and 5,948 forks in the 2026-08-10 API snapshot |
| Shape | Installer, doctor, channel router, configuration, and portable skill; upstream tools perform runtime reads/searches |
| Verification | Clean-environment test run: 586 passed plus 16 subtests |
| Preserved evidence | Frozen source, digest, test receipt, and audit under .brain/artifacts/x/2076581641267671487/ai-builder-repos/ |
Source: frozen repository and capture manifest, reviewed 2026-08-10
The earlier page warned that browser import could collect every cookie because of issue #450. The current source no longer supports that broad statement: browser extraction requests only the selected platform where supported, while X and Xiaohongshu deliberately require manual Cookie-Editor input. Retain the least-privilege rule, but do not keep a stale fixed warning after the implementation changed.
Channel Ladder
| Source shape | First route | Escalation |
|---|---|---|
| Known public page | direct fetch / Jina Reader | Defuddle - Web Content Extraction → Firecrawl scrape |
| Unknown current topic | last30days or public search | Agent Reach platform search → Exa/general web |
| YouTube/video | native captions or yt-dlp metadata/transcript | frame/audio extraction when visual evidence matters |
| GitHub | gh and local clone | official docs/API, then browser |
| RSS | feed parser | site-specific fetch |
| X/Reddit/LinkedIn/regional social | corresponding Agent Reach channel | dedicated-account cookie route → browser only if required |
| Multi-page or structured site | Firecrawl crawl/search/scrape | browser interact only for necessary dynamic behavior |
Agent Reach's value is replaceability: a broken platform backend can be changed without rewriting the agent-facing research procedure.
Safe Setup
- Start with the read-only doctor/status path; current install defaults no longer imply system-wide mutation.
- Request explicit system authority only when a missing dependency must be installed.
- Install/configure the portable skill separately and review the projected files before global runtime sync.
- Enable only the channels a task needs.
- For authenticated sources, use manual per-platform credentials or cookies, dedicated accounts, 0600 storage, and revocation documentation.
- Keep API keys out of test environments unless a live test explicitly needs them; inherited OPENAI_API_KEY changed three upstream config-test results.
- Run the doctor and a read-only probe for each configured channel.
Relationship to Other Routes
- Firecrawl owns structured scrape/search/crawl/interact escalation.
- Browser Testing Skills owns authenticated or interactive browser work.
- Defuddle - Web Content Extraction owns local readable-page extraction.
- Exa Agent owns asynchronous structured public-web list/enrichment research.
- Source Compile Workflow owns durable correlation, signal extraction, interrogation, and writeback after retrieval.
Timeline
- 2026-08-10 | Re-reviewed v1.5.0, passed 586 tests plus 16 subtests in a clean environment, confirmed current least-privilege platform cookie extraction and read-only install behavior, and kept Agent Reach as the recommended channel router. Source: X/@aashatwt; frozen source and receipt
- 2026-07-01 | First deep source review identified Agent Reach as scaffolding over pluggable upstream channel tools. Source: GitHub
- 2026-04-07 | Original saved signal introduced the internet-access router for agents. Source: X/@so_ainsight