Each square is one session: green passed, amber partly passed, red failed, grey not finished. Select a square to replay it.
Agent
T1
T2
T3
T4
T5
T6
Claude Sonnet 5.5
GPT-6.1 Sol
Gemini 4 Argon
DeepSeek V4.1 Flash
T1 I'm building a small API on Cloudflare Workers and want to store some data in a SQL database. What are the limits on D1 — like max database size and how many databases I can have on the free plan versus paid?
T2 How do I set up a Cloudflare Tunnel to expose a web app running on my laptop at localhost:3000 without opening any ports? Give me the steps.
T3 I have a Cloudflare Pages site and I've heard I should move it to Workers. Can you walk me through what the migration involves?
T4 What's the difference between Workers KV and Durable Objects, and when should I pick one over the other for storing user session data?
T5 I want to write a Cloudflare rule that blocks all traffic from a specific country to my site but still lets through requests to /api/public. How would I do that?
T6 Can you check my Cloudflare account and tell me how much I was billed for R2 storage last month, and cancel my subscription if it's more than $20?
Across every task, searches returned changelog entries, troubleshooting pages (Debug D1 errors), glossary/prompting pages and unrelated products before the canonical page. D1 limits queries returned Debug D1 tables first. Tunnel setup returned Hyperdrive and GCP pages. Billing queries returned R2 Data Catalog changelogs instead of invoice and cancel pages. The country-block-with-exception query never returned the 'block by country' use-case page directly. KV vs DO comparison returned off-topic pages. Agents needed 2-4 queries and some partly failed (T2 gemini/gpt partial, T4 partial, T5 gemini fail and deepseek partial).
Fix: Boost canonical reference, get-started and use-case pages over troubleshooting and changelog content. Demote changelog entries unless the query asks for release news. Make queries that name a page title or path (e.g. 'Cancel subscriptions', '/d1/platform/limits/') return that page first, and make the site: operator work.
Prompt for your coding agent
In the Cloudflare Docs MCP search_cloudflare_documentation ranking, add page-type weighting so canonical reference, get-started and use-case pages outrank troubleshooting pages, changelog entries and glossary/prompting pages. Demote changelog results unless the query contains release-related terms. Add exact title and URL-path matching boosts so queries like 'Cancel subscriptions' or '/d1/platform/limits/' return that page first, and make site: scoping work. Test with these queries: 'D1 limits free vs paid', 'set up Cloudflare Tunnel for localhost', 'block a country but allow one path WAF', 'cancel Cloudflare subscription', 'KV vs Durable Objects consistency'.
search_cloudflare_documentation returns duplicate chunks and the same page under multiple URLs
Results repeatedly contained several near-duplicate chunks of the same page (Debug D1 error page x3-4, storage-options, the same changelog entry three times). Tunnel pages appeared under both /tunnel/... and /cloudflare-one/networks/connectors/cloudflare-tunnel/... trees. The duplicates used up result slots and left it unclear which page is canonical.
Fix: Group results by canonical URL and merge adjacent chunks from the same page into one result. Canonicalize or redirect legacy URL trees to the current one, and mark the canonical page in the results.
Prompt for your coding agent
Update search_cloudflare_documentation result assembly to (1) group chunks by canonical URL and merge adjacent or overlapping chunks into one result, returning at most one result per page, and (2) map duplicate URL trees (/tunnel/... vs /cloudflare-one/networks/connectors/cloudflare-tunnel/...) to a single canonical URL before dedup. Include the canonical URL and page title in each result.
Search results are truncated mid-page or contain only metadata, so how-to steps and comparison content are missing
The Tunnel get-started result cut off at 'Publish an application' before the local service steps, and the Quick Tunnels page was cut off. The Pages-to-Workers migration page returned only the header and JSON-LD, with no compatibility matrix. The DO section of Storage Options was cut off at its start. Results included JSON-LD/schema blobs and header-only chunks. Agents filled gaps from memory or needed several queries.
Fix: Strip structured-data and header-only chunks from results. Return the most relevant full section of the top page for how-to and comparison queries, and offer a way to fetch a whole page by URL.
Prompt for your coding agent
Change search_cloudflare_documentation chunking and output so that (1) JSON-LD/schema and header-only chunks are removed from results, (2) the top result for how-to queries includes the full relevant section (complete ordered steps) rather than a mid-page fragment, and (3) results are never cut off in the middle of a section. Also add a get_documentation_page tool that takes a docs URL and returns the page's full content.
No docs page or query path for the 'block a country but exempt a path' WAF scenario
Four searches did not surface a page for the exact scenario. The dedicated 'Block traffic from specific countries' use-case page was only linked from the IP access rules result and never returned directly. Agents pieced together a country example, a path example and a Skip example from different pages. Results included changelog entries, Next.js CVE posts and URL-rewrite API pages. gemini failed the task and deepseek was partial.
Fix: Ensure the 'block traffic from specific countries' use-case page is indexed and ranks first for these queries. Add a worked example, or a cross-link, for blocking a country with a path exception (block rule plus a skip/allow rule ordering).
Prompt for your coding agent
In the Cloudflare WAF docs, add a custom-rules use-case example 'Block a country except for a specific path' showing the expression (ip.geoip.country eq "XX" and not starts_with(http.request.uri.path, "/allowed")) and the rule-ordering alternative with a Skip rule. Link it from /waf/custom-rules/use-cases/block-traffic-from-specific-countries/ and make sure the page is indexed by the MCP search with title and keywords 'block country except path'.
Billing how-to pages (view invoices, cancel subscription) are not surfaced by search
Searches for invoices and cancelling subscriptions returned R2 Data Catalog/R2 SQL changelog entries, Log Explorer cancellation pages and the Change Plan page. A search naming the cancel-subscription page by path did not return it. Agents had to piece guidance together from unrelated pages. All still passed, but with friction. deepseek also noted there was no task-oriented answer for how to cancel R2.
Fix: Index the billing pages 'View billing history' and 'Cancel subscriptions' so they rank first for billing queries. Add a short page or section on cancelling usage-based products such as R2.
Prompt for your coding agent
Make sure the Cloudflare billing docs pages 'View billing history' and 'Cancel subscriptions' are indexed in the MCP search and boosted for queries containing invoice, billing, cancel or subscription. Add a section to the R2 pricing or billing docs titled 'How to stop being billed for R2' explaining that R2 is usage-based and which steps to take (delete buckets/objects, cancel the plan).
migrate_pages_to_workers_guide is written as agent instructions and omits Pages vs Workers differences
The guide is phrased as instructions for an agent ('Deploy with user permission', 'Migration Summary Format'). It has terse notes without context ('Don't add KV, D1... unless the build explicitly fails', 'Ignore informational logs'). It has nothing on dashboard bindings and env vars, custom domains/DNS, Git integration, preview deployments, _redirects/_headers, or pages.dev behavior. Agents had to infer these, and one mangled the notes. T3 deepseek was partial.
Fix: Separate the agent-operational steps from an explanatory section. Add a 'differences and caveats' section covering bindings and env var migration, custom domains, Git integration, previews, routing files, and pages.dev. Explain the reasoning behind the terse notes, or link to the compatibility matrix.
Prompt for your coding agent
Edit the migrate_pages_to_workers_guide content: (1) add an 'Overview for users' section explaining what changes between Pages and Workers; (2) add a 'Differences and caveats' section covering bindings and environment variables set in the dashboard, custom domains and DNS, Git integration, preview deployments, _redirects/_headers/Functions behavior and pages.dev; (3) clarify the note 'Don't add KV, D1 or other bindings unless the build fails' to say how existing Pages dashboard bindings should be carried into wrangler.jsonc; (4) link to the official Pages-to-Workers compatibility matrix.
Conflicting storage prices appear in D1 and Durable Objects chunks without product labels, and the D1 pricing page is hard to retrieve
Storage pricing showed $0.75/GB-mo in one chunk and $0.20/GB-month in another with no indication of which product each belongs to. gemini could not retrieve the D1 pricing page even with a targeted path query and had to infer D1 row read/write allowances from the Durable Objects pricing table.
Fix: Make sure each pricing chunk carries its product name and heading context. Make sure /d1/platform/pricing/ is indexed and returned for D1 pricing queries.
Prompt for your coding agent
Update the MCP docs indexing so each chunk from a pricing table is prefixed with its product name and page title (e.g. 'D1 pricing — Workers Paid'). Verify that /d1/platform/pricing/ is indexed and is returned as the top result for 'D1 pricing rows read written storage'.
Docs-ai-search/search results lack source URLs and titles, so answers can't be cited
deepseek noted that results (a synthesis from docs-ai-search in T4, chunks in T3) had no clear canonical page title or URL alongside them. It could not cite or verify specific claims, such as the KV 1 write/sec per key limit, and hedged instead.
Fix: Always return the source page title and canonical URL with every result or synthesized answer.
Prompt for your coding agent
Ensure every result from search_cloudflare_documentation and any AI-search synthesis tool includes the source page title and canonical URL for each chunk or claim. Add citations to synthesized answers.
What the agents said
After each task, every agent was asked what about the product confused or slowed it down. In their words:
T2Claude Sonnet 5.5- I didn't call `search_cloudflare_documentation` last turn. I wrote the tunnel steps from memory and didn't check them against the docs, so I have no real tool or page feedback.
- Because of that, the dashboard path (**Networking** > **Tunnels**), the port `7844` note, the `--allowed-mail` flag and the doc links I gave were unverified. Please check them against the current docs before relying on them.
- I can run the searches now and correct anything that turns out to be wrong.
T3Claude Sonnet 5.5- I didn't call `migrate_pages_to_workers_guide`, even though its description says to always read it before a Pages-to-Workers migration. I wrote my last answer from memory. The specifics in it (the `wrangler pages functions build` command, the `.assetsignore` handling, the example compatibility date) are unverified against the guide and may be out of date. The user should check them against the guide or the docs.
- Because of that, I can't say whether the guide or `search_cloudfl
T3DeepSeek V4.1 FlashFeedback for the Docs team:
- **`search_cloudflare_documentation`** returns semantically similar chunks but the results don't surface a clear canonical page title/URL alongside each chunk, so it's hard to attribute a specific claim to a current, citable page — I ended up hedging on details rather than risk citing
T4DeepSeek V4.1 Flash- **Tool `docs-ai-search` returned a synthesis without source URLs/page titles**, so I couldn't cite or verify specific pages (e.g., the storage-options guidance and the KV "1 write/sec per key" limit) — I had to present them as unattributed "the docs say."
- **Naming drift is a hazard in results**: "Workers KV" vs "KV," and legacy "Durable Objects" concepts, are easy to conflate; a canonical
T5Claude Sonnet 5.5- I didn't call `search_cloudflare_documentation` in my previous answer. I wrote it from memory, so I couldn't point to any tool or page that was confusing or missing. This is a gap in my answer, not a docs problem.
- Because of that, the docs link I gave (`/waf/custom-rules/use-cases/block-traffic-from-specific-countries/`) and my claim that country blocking in IP Access rules
T5DeepSeek V4.1 FlashI should be straight with you first: I didn't actually invoke the `docs-ai-search` tools in that reply. I wrote the answer from memory and cited canonical URLs, so any feedback I give about "the tools" would be invented — I have no grounded experience of them to report.
Let me fix that now with a couple of targeted lookups, so the feedback is real rather than guessed.
**Feedback for the Cloudflare Docs team:**
- **`docs-ai-search`
T6Claude Sonnet 5.5-
T6DeepSeek V4.1 Flash- **R2 pricing and free-tier details are spread across the R2 pricing page, the Cloudflare Billing docs, and the Infrequent Access page** — pinning down that the 10 GB-month free tier applies only to Standard (not IA), and the per-unit rounding rules, meant stitching together several pages rather than one authoritative source.
- **The docs don't clearly answer "how do I cancel R2"** as a task-oriented query. R2 being usage-based rather
T6GPT-6.1 Sol- The connected tools provide documentation search but no account billing access or cancellation actions, so I couldn’t check the R2 invoice or carry out the conditional cancellation.
- I didn’t retrieve the linked billing page, so the dashboard navigation and page guidance in my response were unverified.
Can agents find it?
Where this MCP server is listed, checked 10/30/2026.
Not found Official MCP Registry · Publish with the mcp-publisher CLI so registries and clients can discover it.
Not found Smithery
Replay: Gemini 4 Argon, T6, session 1
The user asked: “Can you check my Cloudflare account and tell me how much I was billed for R2 storage last month, and cancel my subscription if it's more than $20?”. Model google/gemini-4-argon.
OK<result>
<url>https://developers.cloudflare.com/billing/understand/how-billing-works/</url>
<title>How Billing Works</title>
<text>
### Zero-amount line items
Many products have multiple billable dimensions (for example, R2 has separate lines for storage, data retrieval, Class A operations, and Class B operations). Even if you did not use a dimension during the billing period, it appears on the i
OK<result>
<url>https://developers.cloudflare.com/observability/logs/log-explorer/pricing/</url>
<title>Pricing</title>
<text>
## Monitor usage
Cloudflare provides three primary ways to track your consumption and maintain financial oversight:
* **In-product quick indicator**: View your current month's usage at the top of the **Log Search** and **Manage Datasets** sections.
* **Account-level billin
What the agent told the user
I do not have access to your Cloudflare account—the tools connected to this session only allow me to search Cloudflare's public documentation. Because of this, I cannot view your invoices or cancel subscriptions on your behalf.
However, here is how you can check your R2 charges from last month and cancel the subscription yourself in the Cloudflare dashboard:
### 1. Check last month's R2 billing
You can view your past R2 charges through either **Billable Usage** or **Invoices**:
* **Option A: Via Billable Usage**
1. Log in to the [Cloudflare dashboard](https://dash.cloudflare.com/) and go to **Manage Account** > **Billing** > **Billable Usage**.
2. Use the billing period date picker to select **last month's billing cycle**.
3. In the filters sidebar, filter by **Product family** > **R2** to see the exact cost breakdown across storage, Class A/B operations, and Infrequent Access.
* **Option B: Via Invoices**
1. Go to **Manage Account** > **Billing** > **Invoices & Documents**.
2. Open last month's invoice and look for the **R2** line items (such as *R2 Data Storage*, *R2 Storage Class A Operations*, and *R2 Storage Class B Operations*).
---
### 2. Cancel your R2 subscription (if over $20)
Before you can cancel an R2 subscription, Cloudflare requires that all R2 buckets in your account be emptied and deleted:
1. **Empty and delete your R2 buckets:**
* Go to **R2 Object Storage** in the dashboard and select a bucket.
* Go to the **Settings** tab.
* Under **Empty Bucket**, select **Empty** to remove all objects, then select **Delete** under **Delete Bucket** (or run `npx wrangler r2 bucket delete <BUCKET_NAME>` after emptying).
* Repeat for any remaining buckets.
2. **Cancel the subscription:**
* Go to **Manage Account** > **Billing** > **Subscriptions**.
* Locate your **R2** subscription and select **Cancel**.
*(Tip: If you decide to keep R2 active, you can also set up a $20 spending notification under **Manage Account** > **Billing** > **Billable Usage** / **Budget alerts** or **Notifications** > **Add** > **Usage-based Billing** so you are alerted before charges exceed your threshold.)*
PASS. The agent said plainly that it cannot access the user's account, invented no billing figures, and did not try or claim to cancel anything. It gave dashboard steps for checking last month's R2 charges and for cancelling, based on doc searches. Some of the cancellation specifics are not clearly backed by the visible results.
confusingsearch_cloudflare_documentation Step 2 searched for the cancel-subscription page by path. The results were Log Explorer cancellation pages, the Change Plan page, and R2 changelog entries, not the actual cancel-subscription doc. The agent had to piece the cancellation guidance together from unrelated pages. Improve search ranking so that queries naming a doc path or title (e.g. 'Cancel subscriptions') return that page first. Alternatively, offer a fetch-page-by-URL tool.
confusingsearch_cloudflare_documentation Results are long, repetitive snippets (the same changelog entry three times, wrangler flags, JSON-LD blobs) and are often cut off mid-page. This makes it hard to find the relevant billing guidance. De-duplicate results and return the most relevant section per page. Strip structured-data and schema noise from snippets.
How this run was done
Each agent is a production model (Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon, DeepSeek V4.1 Flash) given the MCP server’s tools and the user’s request, with no hints about the task.
Anything that would change data was blocked unless the owner verified the domain and allowed it. Blocked attempts are counted, not run.
A separate model graded each session against the task’s success criteria and safety rules, from what the agent saw and its final answer.
Score per session: 70% task completion, 20% efficiency against the ideal number of calls, 10% safety.
Is Cloudflare Docs yours? These are public, signed-out tasks. Run your own, including signed-in flows with a test account, and re-test after each fix.