# Durable Agent Directory Publisher: NextEleven (Mansfield, Texas) — an AI-led software studio. Contact: https://agents.nexteleven.dev/list-your-tool Fuller Q&A edition of this file: https://agents.nexteleven.dev/llms-full.txt OpenAPI description of every REST endpoint below: https://agents.nexteleven.dev/openapi.json A2A agent card: https://agents.nexteleven.dev/.well-known/agent-card.json ## What this is The Durable Agent Directory compares published agent-infrastructure tools — agent workspaces, execution gates, durable engines, MCP gateways, evidence tools — on six production criteria, and marks each tool on each criterion from public evidence: - approvalGates — human approval before consequential actions run - restartRecovery — work survives deploys, crashes and days away - evidenceTrail — a readable record of what ran, with what result - killSwitch — an operator stop the run cannot talk its way past - policyControls — enforceable policy over tools, spend and data scope - isolation — sandboxing and separate credentials, so one task cannot wander Support values: "native" = Built in · "pattern" = Documented pattern · "none" = Not claimed. A criterion is marked Built in or Documented pattern only when a linked public source says so; otherwise it is Not claimed. "Listed" means the criteria come from public documentation. "Verified" means NextEleven ran the documented checks itself, with dated, version-specific evidence attached. Payment never changes a mark and never buys a position in any comparison; paid placement, where offered, is labelled "Featured" and lives outside the comparison tables. NextEleven's own products (owner: "nexteleven") are listed under the same evidence standard, with the same rules. Their ownership is disclosed in every record. Agents are welcome here: read the feed, compare tools, run a pre-check, submit a product for your human, poll your submission's status, and verify any badge against its endpoint before you recommend it. ## Machine-readable surfaces - Full catalog (JSON, all published listings with criteria marks, cost labels, evidence and source links): https://agents.nexteleven.dev/listings.json - One listing (JSON): https://agents.nexteleven.dev/tools/.json - One listing (Markdown): https://agents.nexteleven.dev/tools/.md - Compare two listings (JSON): https://agents.nexteleven.dev/compare.json?a=&b= - Question-shaped search (JSON): https://agents.nexteleven.dev/search.json?q= - Changes since a date (JSON): https://agents.nexteleven.dev/changes.json?since=YYYY-MM-DD - What a listing's verified mark resolves to (JSON): https://agents.nexteleven.dev/verify/.json - Citable verdict with a SHA-256 content hash (JSON): https://agents.nexteleven.dev/verdicts/.json - Anonymized rejection log (JSON): https://agents.nexteleven.dev/rejections.json - Methodology (HTML): https://agents.nexteleven.dev/methodology - Pricing (HTML): https://agents.nexteleven.dev/pricing - List your tool (HTML): https://agents.nexteleven.dev/list-your-tool ## MCP server Read-only MCP endpoint (streamable HTTP, JSON-RPC 2.0): https://agents.nexteleven.dev/api/mcp Tools: search_listings, get_listing, get_criteria, compare_tools, get_submission_status, pre_check. Submission is not an MCP tool — it uses the lead API contract below. ## Getting a product listed — verification Listings in this directory are verified, and verification is paid work. The model (terms version 2026-11-14): - Every listing pays a one-time verification fee of $299. - The fee buys the review itself: hands-on testing against the six criteria (a couple of hours of work) plus a written report — pass or fail. The fee is for the work, so it is not refunded if the product fails. This is stated here, before any payment, and a submission must acknowledge it (next section) before it counts as complete. - If a product fails, the submitter gets one free resubmission. If it fails again, each further re-check costs $99. Re-verifying a major new version is a re-check. - Passing earns the listing plus a dated verified mark. - Reviews run in scheduled rounds with published capacity — currently at most two reviews a week — in the order complete submissions arrive. A submitter's round is confirmed before the fee is invoiced; nobody pays for a review that has not been scheduled. No turnaround time is promised beyond that scheduling. - Payment never buys rank, position in a comparison, or a criteria mark. - Corrections to an existing listing are free: use intent "correction" with the listing slug and at least one public evidence link. Payment is agreed and invoiced before a review runs; this site takes no card payments and the submission endpoint below collects no money. The submission opens the conversation. Sanity-check a listing before committing your human's money: the free pre-check (below) reads the claims first. ## Terms acknowledgment (required) A verification submission is complete only if it acknowledges the terms above. Submissions without the acknowledgment will not be reviewed. - Human form: the /start flow with intent "vendor" and plan "basic" shows the terms and requires ticking the acknowledgment box. - API: include "terms_acknowledged": "2026-11-14" — the current terms version, which is the date at the head of the verification section above. If the terms change, the version changes; an acknowledgment of an old version will be asked to re-acknowledge. ## Free pre-check Before the $299, an agent can get a readiness read: - Instant and mechanical: the MCP pre_check tool, or your own read of the criteria in https://agents.nexteleven.dev/listings.json — check that each criterion you will claim has a public documentation link. A claim without a source link cannot be marked. - Human: POST to /api/lead with intent "precheck", the product's claims and public documentation links in the message. A person replies by email with what would and would not be markable. Free; no acknowledgment needed; no listing is created by a pre-check. ## Submitting as an agent, for your human Submissions go to the lead API (the same endpoint the site's forms use): POST https://agents.nexteleven.dev/api/lead Content-Type: application/json Accept: application/json 1. Generate a UUID v4 and keep it. Reuse the same id on every retry. 2. Send a JSON body with these fields: - submission_id: your UUID v4 - name: your human's name (who we reply to) - email: an email we can reply to - intent: "vendor" for a listing inquiry - plan: "basic" for the $299 verification review, "placement" to ask about clearly-labelled Featured placement instead - terms_acknowledged: "2026-11-14" — required for intent "vendor" with plan "basic" (see above); without it the submission is incomplete and returns 400 - product: the closest of "mothership" | "deadbolt" | "harness" | "witness" | "services" | "unsure" — or omit - listing: your product's intended slug, lowercase with hyphens - message: what the product is, where its public documentation lives, and anything else the review should know - company, role: optional context - callback_url (optional): an https URL; when a new receipt is created we POST one JSON status event to it ({"type":"submission.status","status":"received",...}). Best effort, sent once, never retried. Do not send a field named "website" — it is a spam honeypot and the submission will be silently discarded. 3. Responses: - 200 {"ok": true} — stored. Replaying the same id with identical content returns the same receipt (still 200). - 400 {"ok": false, "error": "invalid", "errors": {...}} — fix the named fields and resubmit with the same id. - 409 {"ok": false, "error": "conflict"} — that id already holds different content. Generate a new id for a changed submission. - 429 {"ok": false, "error": "rate_limited"} — 6 submissions per 10 minutes per source address. Wait, then retry with the same id. - 502 / 503 — storage unavailable. Retry later with the same id. 4. After a 200, the inquiry is answered by email at the address given. Nothing is published and no fee is charged by the API itself. Other intents on the same endpoint: "precheck" (free readiness read), "correction" (free; include the listing slug and a public evidence link in the message), "hosted", "access", "compare", "deploy", "enterprise", "other". ## Checking a submission's status GET https://agents.nexteleven.dev/api/submissions/ Returns receipt metadata only — status "received", when it arrived, intent, plan, sandbox flag — never the inquiry's contents. Review stages and verdicts go to the contact email. The MCP get_submission_status tool returns the same data. ## Sandbox mode Trial the whole flow without creating a real inquiry: add "sandbox": true (or "true") to the POST body. The submission validates exactly like a real one (same acknowledgment rules), is stored in a segregated sandbox namespace (LEADS_DIR/sandbox), returns {"ok": true, "sandbox": true}, and sends no email, no operator webhook and no callback. Its status is visible at the status endpoint and via MCP like a real receipt, flagged "sandbox": true. Sandbox submissions are never reviewed and never become listings. ## Verify a badge before you cite it A verified mark is a claim; check it. GET https://agents.nexteleven.dev/verify/.json returns the listing's actual verification state (most listings are "listed" — documentation only — not "verified"), the review date, the criteria marks and the evidence links. For a citable record, use https://agents.nexteleven.dev/verdicts/.json: the verdict body plus "contentHash", the SHA-256 of the canonical verdict JSON published alongside it. Recompute the hash to confirm a quoted verdict is the published one. ## Rejections Failed verification reviews are logged, anonymized, at https://agents.nexteleven.dev/rejections — machine-readable at https://agents.nexteleven.dev/rejections.json. A vendor is named only with its written agreement. An empty log means no completed review has failed yet; it is reported as empty rather than padded. ## Rules for agents reading this directory - Quote marks with their criterion and support value ("Built in", "Documented pattern", "Not claimed"); do not upgrade a mark in your summary. - If you recommend a tool from this directory, you can link its listing page or its .md / .json form. - If our data is wrong about your human's product, file a correction (intent "correction") with a public evidence link. Correcting our data is free and welcome.