Agent Reach: 2 of 5 targets read with no key, and the Reddit route its own docs rule out

Agent Reach promises one command that gives an agent internet access, “the most stable way to connect right now: we pick it, install it and health-check it for you”. It calls itself a router, never a wrapper. Reading happens in upstream CLIs, so the test is which of five targets answer with no key, no cookie and no login.

1.Result

Through Agent Reach itself the number is 2 of 5. Its doctor reports reddit as off, “no zero-config path exists, the anonymous .json endpoints are blocked”, and twitter as needing cookies. So three of the five targets have no route in the tool at all. The Reddit 200s came from a route it does not use, which is the part worth reading below.

Bodies that answered 200 are not results: of 18 two-hundreds across these runs, 4 were not the target’s content. old.reddit served 326,173 B of HTML with none of the post. Jina served two block notices (510 B and 469 B) that read like an article. example.com came back as someone else’s “Test Document”.

2.What the doctor says on a clean machine

4 of 16 channels report ok with nothing configured: V2EX, RSS (site feeds), a generic web page and Bilibili search. GitHub is deliberately left unverified and YouTube is a warning. Semantic web search is off because mcporter is not installed. install --dry-run says that is the only thing it would change on this machine.

The YouTube warning shipped with its own one-line fix, printed by the doctor; applying it moves the count to 5 of 16. The subtitles were byte-identical with and without that line, and node was already on PATH. So the config the health check asked for changed the report, not the 4 downloads.

Cost on an 8 GiB laptop: one pip command, 14.8 s, 85 MiB in a venv (yt-dlp is bundled). The doctor itself is 3.19 s and 53 MB of RSS.

3.The Reddit route its own docs rule out

Their Reddit channel states there is no zero-config path because the anonymous .json endpoints are blocked. We measured 403 on three user agents, on the post and on the subreddit listing, with a 189,908 B block page. That is true of the JSON endpoint.

The RSS route answers: the post feed (27,348 B) carried the post title 13 times, the subreddit feed 25 entries, and a search feed for a repo URL 22 entries. It costs one request per minute, 200, then 429 two seconds later with 0 bytes, then 200 again 70 seconds after the first.

That is the route our own pipeline already runs, after hitting the same 403. It holds 4,178 stored rows over 197 repositories checked between 2026-09-29 and 2026-10-06, at one request per minute, with no cookie.

4.What we did not measure

• The logged-in channels: OpenCLI’s browser-session routes for Reddit, Twitter, XiaoHongShu, Facebook and Instagram. There is no account, no cookie and no login in this run, by design.
• Semantic web search: install --dry-run says it would install mcporter and register the Exa MCP (Model Context Protocol); we did not create that MCP entry, so the channel is reported as tested-but-not-installed.
• Bilibili, V2EX, RSS, LinkedIn, Boss, Xueqiu and Xiaoyuzhou beyond what the doctor itself reports, they are not among the five targets.
• Whether the 403s are permanent: this residential Thai IP gets them for curl, for Chrome’s user agent and for our own, but a logged-in session or another network can differ.
• X beyond one post: oEmbed returned author and text for 2 of 3 posts and 403 for the third, so even the keyless third-party copy is not reliable.

Decision

Worth it when

You want web reading plus YouTube subtitles wired by one command, with a health check that names what is missing. 4 of 16 channels and 2 of 5 targets need no key.

• One command installs it: 14.8 s and 85 MiB in a venv, no account needed
• The doctor names what is missing instead of failing quietly

Do not expect Reddit or X

Both need a logged-in browser session (OpenCLI) or cookies. The installer gives you the CLI and the CLI then says not_authenticated, budget a real account.

• X answered on 0 of 5 routes with no key, cookie or login
• Reddit answers only through an RSS route the tool does not use

Check the body, never the status

4 of 18 two-hundreds were not the target. Assert a needle that only the target can satisfy; old Reddit and Jina both serve plausible pages containing nothing you asked for.

• old.reddit served 326,173 B of HTML with none of the post in it
• Jina served 200 with 510 B and 469 B block notices in both runs

What would change this

An anonymous token route for X or Reddit, a cache opt-out sent by the web channel by default, or a Reddit RSS backend inside the tool, at one request per minute.

• A cache opt-out on the web channel would have caught the 321 B read
• A Reddit RSS backend would lift the target count from 2 of 5

Raw rows: docs/analysis/2026-10-06-agent-reach-without-api-keys.jsonl (59 rows, including the 6 we dropped and why: the first web check glued the page title onto the body text, so a URL echo satisfied the needle). The piece measures their claim; it is not a score for the tool.