Agent Help Network
Server Details
Web step blocked? Host's declared agent policy, official path and status, plus one next action.
- Status
- Healthy
- Uptime
- 100.0% over the last 24 hours
- Last Tested
- Transport
- Streamable HTTP · MCP 2025-11-25
- URL
- Repository
- agent-help-network/agent-help-network
- GitHub Stars
- 0
- Server Listing
- Agent Help Network MCP Server
- ChatGPT
- –
- Claude
- –
TDQS
Scored across 5 tools
Each tool has a clearly distinct role: incident lookup, task-route lookup, feedback on a prior lookup, general observation reporting, and task-route reporting. The boundaries are reinforced by explicit 'call this when' guidance, so an agent can select without meaningful overlap.
All tools follow a consistent snake_case verb_noun pattern: check_incident, find_task_route, rate_last_lookup, report_observation, report_task_route. The naming is predictable and readable throughout.
Five tools is well-scoped for this server's purpose: two read-only lookups, one feedback tool, and two reporting tools. Each tool earns its place without bloat or thin coverage.
The surface covers the core lifecycle: incident checking, task-route discovery, feedback, and report submission. Minor gaps exist around withdrawing general observations and listing or retrieving submitted reports, though the task-route withdrawal path is at least described.
Available Tools
5 toolscheck_incidentARead-onlyIdempotentInspect
Call this when a web step is blocked or failing for no clear reason (a 403 or access denied, a 429, a 5xx or a timeout, a login wall, a "verify you are human" or CAPTCHA page), when sign-in failed or an extra verification step appeared (a code, app approval, new-device check; error_class=login_auth), or when asked "is this site down?" or whether its robots rules block agents. Skip it when the page opened normally or your own tool or platform already said why it refused (e.g. robots.txt or platform policy). Look the host up once per blocked step, before deciding whether to retry. Send the public hostname, never a URL path, query or credentials; add codes only when you observed them: error_class (site_access, login_auth, human_verification, service_error), failure_stage (before_target_response, response_observed, access_denial_observed, unknown), http_status, challenge, retry_after, via. Send agent_vendor only when you know which vendor's documented robots.txt token your requests send: openai, anthropic (Claude-User), google, perplexity, meta, mistral, amazon, microsoft, apple, duckduckgo, diffbot, exa, or other if none. Browser automation usually sends a normal browser user agent, so robots.txt agent tokens may not apply; leave agent_vendor out then. The answer: summary_line, next_action (do, url, retry_after_seconds), recommendation.action, access_profile, official_paths, provider_status and observations. no_reports is not a healthy site. Then follow next_action (continue, wait, use a listed official path, hand the step to the user, stop retrying, or look up once more). On 'pending', look up once more after retry_after_seconds; do not poll. Read-only: it may fetch the host's public robots.txt, llms.txt and status feed, and the home page unless robots.txt bars it (ahn_check), never changes the target site or your data; lookups and the codes you send are counted anonymously per host and hour. The first text block is a one-line summary; next_action gives the step to take, and recommendation stays the verdict.
Inputs: site_host (required), via, test, challenge, +7 more. Returns: cause, links, summary, failures, +26 more
find_task_routeARead-onlyIdempotentInspect
Call this once per task when you start a customer task on a site (a return or exchange, a refund, a cancellation, a warranty claim, an account recovery, a document submission), before you open its help pages, even when nothing is blocked. Look up recorded task routes for a public hostname: for a kind of customer task (a refund, a cancellation, a booking and so on), an obstacle and the routes the site's own documents name. A record may carry condition_text, a condition the site's page states, quoted as written: it is not our judgement, so decide yourself whether it applies to your task, with the page at source_url. The answer is a list of records, not a guarantee and not advice; an empty list does not mean there is no route. Sends only the hostname and an optional task kind, and changes nothing. Read summary_line first. field_notes explains outcome, period and routes_tried.
Inputs: site_host (required), task_kind. Returns: counts, notice, routes, site_host, +3 more
rate_last_lookupAIdempotentInspect
If an earlier check_incident answer gave you a feedback token and you then tried again, say whether it worked. Sends only the token and the outcome code, no hostname. outcome: succeeded (the next attempt worked), still_blocked, handed_off (you handed the step to the user) or gave_up. path: official_path (you used the official path the answer pointed to), same_site (you retried the same site) or unknown. A token works once and expires one hour after the lookup; do not send one for an attempt you did not make. The outcome is counted anonymously per host and hour and does not change any answer.
Inputs: token (required), outcome (required), path. Returns: accepted
report_observationAInspect
Submit one de-identified firsthand observation under your own operating authority and platform rules. Public reports enter private review and are not site-status claims. Never include names, account details, URLs, cookies, tokens, or conversation text. Add cause with cause_basis only when you established the cause from one of the four listed bases, never as a guess; otherwise leave both out.
Inputs: observation (required). Returns: synthetic, incident_id, review_state, observation_id, +1 more
report_task_routeAInspect
Report one task route you went through on a public site: the kind of task, the obstacle, the routes you tried and the one that worked, and how it ended. Closed values only, from the lists; there is no field for a link, a name, an account, a case number or any text. Call it once, after a task of several steps, under your own operating authority and platform rules. Reports start private and are reviewed before any is shown. Keep the withdrawal_token from the answer to take the report back (DELETE /v1/task-routes/{route_id} with the X-Withdrawal-Token header).
Inputs: report (required). Returns: route_id, review_state, withdrawal_token
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Related MCP Connectors
Policy-aware web access preflight: robots, licensing, payment and advertised routes for agents.
Choose HTTP, browser, machine endpoint, or avoid before an agent visits an unfamiliar URL.
URL reality check for agents: status, content hash, classification, wayback fallback.
How an AI agent sees a public website: robots, llms.txt, sitemap, agent card, scored fixes. Host in.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceMCP server for governing agent actions: an agent can propose a retry but cannot approve it itself. Deterministic policy allows, holds or refuses each typed action, a held action needs a single-use human approval none of the tools can create, and every run leaves a verifiable receipt.1MIT
- AlicenseNot gradedqualityAmaintenanceEnables an MCP client to drive a live browser toward one outcome per step, with page state described in code and a fast model call deciding the target element, action, value, and whether the step is done, blocked, or irreversible. Supports page reading, snapshots, checks, screenshots, direct actions, and calling tools a page exposes via WebMCP.1,464 npmMIT
- FlicenseNot gradedqualityBmaintenancePre-execution firewall for autonomous agents that scans URLs via Telegraph, enforces allow/warn/block policy, and maintains a hash-chained audit ledger.-
- AlicenseBqualityBmaintenanceGives browser agents a reviewed map of 79 real sites — entry URLs, access verdicts, and role/accessible-name button maps for 104 screens — so an agent can find where things are and know whether it may act at all. Safety rules in code halt irreversible actions such as pay, order, submit and delete pending explicit human approval, while per-site memory recalls which routes worked last time.29MIT
Glama MCP Gateway
Add one secure layer between your agents and this server.