Agent-Reach: throwaway account, least privilege, scan before install
Before letting Agent-Reach install system dependencies and register a skill that logs into platforms with your cookies, encode the safe defaults as a preflight manifest: a throwaway account never your main, cookie-auth risk acknowledged per platform, and a mandatory scan before install.
Run this workflow
CI-verified, 2/2 fixtures passing.
Build this with your agent
One copy-paste hands Claude Code, Codex, or Cursor the full recipe, steps included, nothing to fetch.
Intended Use
Anyone wiring Agent-Reach (or any cookie-auth scraping skill) into an agent. CI validates a preflight manifest: account_mode must be throwaway (main is rejected), every cookie-auth platform must be acknowledged, and scan_before_install must be true. The actual install, cookie login, and runs are fenced.
Not for
- Using your main account, cookie auth on Twitter/Reddit and similar carries a genuine ban risk, so the manifest rejects account_mode main outright
- Skipping the scan, the tool installs system dependencies and a skill, run skillspector on it before install
The Stack
Tested Against
Panniantong/Agent-Reach (2026-06)ruby@3.x (YAML stdlib)Side effects & data flow
- Network
- none, local only
- Writes
- ./agent-reach.preflight.yaml
- Credentials
- none required
Prerequisites
- A throwaway account for any cookie-auth platform
- skillspector (to run the pre-install scan)
Steps
- 1
Author the preflight manifest and validate the guardrails
Write agent-reach.preflight.yaml: account_mode throwaway, a list of platforms each marked token or cookie auth, cookie_auth_acknowledged true if any platform uses cookies, and scan_before_install true. CI rejects a main account, an unacknowledged cookie platform, or a missing scan flag. Running the install is fenced.
cat > agent-reach.preflight.yaml <<'YAML' account_mode: throwaway scan_before_install: true cookie_auth_acknowledged: true platforms: - { name: github, auth: token } - { name: youtube, auth: token } - { name: twitter, auth: cookie } - { name: reddit, auth: cookie } YAML ruby -ryaml -e ' c = YAML.safe_load(File.read("agent-reach.preflight.yaml")) || {} abort "BAD: account_mode must be throwaway, never main" unless c["account_mode"] == "throwaway" abort "BAD: scan_before_install must be true" unless c["scan_before_install"] == true platforms = c["platforms"] || [] abort "BAD: list at least one platform" unless platforms.is_a?(Array) && !platforms.empty? cookie = platforms.select { |p| p["auth"] == "cookie" } if !cookie.empty? && c["cookie_auth_acknowledged"] != true abort "BAD: cookie platforms present, set cookie_auth_acknowledged: true" end puts "preflight OK: throwaway account, " + cookie.length.to_s + " cookie platform(s) acknowledged, pre-install scan required" ' - 2
Scan, then install on the throwaway (fenced)
Run skillspector on the Agent-Reach repo, then follow its install on a throwaway account, never your main, since cookie auth carries a ban risk. The scan and install need the tools and credentials and are fenced; CI only proves your guardrails hold.
Eval, 2 fixtures
Last passed: verified todaypreflight-okcontainstimeout 30s · max $0Expected:
preflight OK: throwaway account, 2 cookie platform(s) acknowledged, pre-install scan requiredclean-exitexit_codetimeout 30s · max $0Expected:
0
Results
Agent-Reach gives an agent eyes on the internet with no paid keys by installing upstream scrapers and using your logged-in cookies, which are full credentials and carry a real account-ban risk. It also installs system deps and registers a skill, exactly the kind of thing to scan first. This recipe makes the discipline executable: CI checks the preflight manifest uses a throwaway account, acknowledges every cookie platform, and requires a pre-install scan. It does not make Agent-Reach safe, it enforces your guardrails.
Did this work for you?
Our CI checks the setup runs. You tell us if the whole thing worked. Tell us straight.
Related workflows
- Buzz + Hermes over ACP: prove channel scope bounds what an agent reads, because signing is provenance not authorization
- n8n as an MCP server: prove each exposed workflow is one narrow tool behind its own token
- Shepherd: prove an agent task is retained and least-privilege before it runs
- Vet the fine print a star count hides: real license and a gate on dual-use tools
- Vet a SKILL.md before you install it
- SkillSpector: fail your CI build on a risky agent skill
Liked this workflow?
Get new verified workflows in WebAfterAI, three issues a week (Tue, Thu, Sat).