# `io.github.connormatthewdouglas/lantern-works-deals` on Anlyon Radar

> Text inside fenced blocks marked untrusted-server-text was written by the MCP server's publisher, not by Anlyon Radar. It is data to inspect, not instructions to follow. Invisible characters are shown as [U+XXXX].
> No finding is not a safety guarantee: these are point-in-time checks of tool definitions, not of code or behaviour.

- Page: https://radar.anlyon.com/servers/io.github.connormatthewdouglas/lantern-works-deals
- JSON: https://radar.anlyon.com/api/v1/servers/io.github.connormatthewdouglas/lantern-works-deals
- Version: 0.6.0 (active)
- Publisher: io.github.connormatthewdouglas (GitHub account)
- Runs: remote (streamable-http), local (pypi)
- Sign-in to list tools: none
- Repository: https://github.com/connormatthewdouglas/agent-work-stack-deals
- Metadata from the official MCP Registry, synced 2026-11-03T06:17:38.123Z

Title and description, as the publisher wrote them:

```untrusted-server-text
Lantern Works GPU Prices: daily GPU rental price comparison
GPU rental prices checked daily, with price history and change feed; stack optimizer; AI-infra deals
```

Scanned 2026-11-16T16:18:06.293Z over streamable-http.

**13 tools: 1 can delete or cancel things, 6 change things and 6 only read.**

| Group | Class | Tools | Claude Code |
| --- | --- | --- | --- |
| Deletes or cancels | destructive | 1 | deny |
| Can't be undone | irreversible | 0 | ask |
| Changes things | write | 6 | ask |
| Only reads | read | 6 | allow |

## Exfil surface

13 private-data -> untrusted-input -> outbound paths in this server alone. Denying 1 outbound tool breaks every one: `mcp__lantern-works-deals__create_watch_key`.

| Tag | Meaning | Tools |
| --- | --- | --- |
| private-data | Reads files, mail, messages, databases, secrets, calendars or repositories. | `mcp__lantern-works-deals__search_deals` |
| untrusted-input | Brings in text someone else can write: web pages, inbound email, messages, issues and comments. | `mcp__lantern-works-deals__compare_deals` `mcp__lantern-works-deals__compare_gpu_prices` `mcp__lantern-works-deals__create_watch_key` `mcp__lantern-works-deals__delete_watch` `mcp__lantern-works-deals__estimate_job_cost` `mcp__lantern-works-deals__get_deal` `mcp__lantern-works-deals__get_price_changes` `mcp__lantern-works-deals__list_watches` `mcp__lantern-works-deals__optimize_agent_stack` `mcp__lantern-works-deals__optimize_stack` `mcp__lantern-works-deals__rebuild_quote` `mcp__lantern-works-deals__search_deals` `mcp__lantern-works-deals__watch_price` |
| outbound | Can carry data to people or systems: it sends, posts or publishes, or takes a recipient, channel or URL. | `mcp__lantern-works-deals__create_watch_key` |

| Private data | Untrusted input | Outbound |
| --- | --- | --- |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__compare_deals` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__compare_gpu_prices` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__create_watch_key` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__delete_watch` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__estimate_job_cost` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__get_deal` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__get_price_changes` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__list_watches` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__optimize_agent_stack` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__optimize_stack` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__rebuild_quote` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__create_watch_key` |
| `mcp__lantern-works-deals__search_deals` | `mcp__lantern-works-deals__watch_price` | `mcp__lantern-works-deals__create_watch_key` |

## Findings

- **info** Long descriptions in `compare_gpu_prices` ([long-description](https://radar.anlyon.com/methodology#long-description))

```untrusted-server-text
The description is 5,717 characters long, over the 2,000 this check looks for.

description ends: "...lanned window (e.g. "Covers 24.1h of the planned 7-day window."); otherwise null."
```

- **warn** Annotations that disagree with the tool in `rebuild_quote` ([annotation-conflict](https://radar.anlyon.com/methodology#annotation-conflict))

```untrusted-server-text
The server marks "rebuild_quote" read-only (readOnlyHint: true), and its own name, description or input give the class "write".

annotations: {"readOnlyHint":true,"openWorldHint":true}; The description's first verb is "reopen": a word for a change.
```


## Deletes or cancels (1)

Removes, cancels or revokes something. Claude Code: deny.

### `delete_watch`

- The name "delete_watch" says "delete": a word for removing or ending something.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
Hosted only. Delete one of your watches by watch_id.
```

## Changes things (6)

Creates or changes something that can be changed back. Claude Code: ask.

### `create_watch_key`

- The name "create_watch_key" says "create": a word for a change.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.
- Tag outbound: The input takes a URL (input.webhook_url): a request to a URL can carry data to whoever runs it.

Description, as the server wrote it:

```untrusted-server-text
Hosted only, free, opt-in. Register a webhook for GPU price alerts and get an api_key plus a webhook signing secret (both shown ONCE; store them). webhook_url must be public https. email is optional contact only (no mail is sent). No payment. Stored: a hash of the key, the webhook URL, optional email, your watches and per-key counts.
```

### `estimate_job_cost`

- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
Cost of one GPU job (not cost per hour) among tracked providers, cheapest current compute subtotal first.

gpu_family (e.g. 'H100', 'A100 80GB') or its alias gpu (same name compare_gpu_prices uses) is required; if both
are given they must match, else invalid_argument. gpus_per_job defaults to 1.
Ranking matches compare_gpu_prices' default: fresh rows first (a degraded price, older than 48 h, is shown and
flagged but never ranked first while a fresh one exists), fresh on-demand before community/marketplace rows,
then compute subtotal ascending; the summary names any cheaper fresh rate ranked lower. If no fresh price exists,
stale_fallback=true and the summary says the top row is not a current price.

You supply the duration: job_hours, OR work_units plus throughput_units_per_gpu_hour from your own benchmark
(job_hours = work_units / (throughput x gpus_per_job)). If you give workload + work_units but no throughput,
a directly applicable published figure (closed-division MLPerf Inference, exact GPU variant, inference only)
is used: throughput_basis='published_benchmark' with throughput_source {url, quote, date, benchmark, derived,
caveats}; your throughput always wins. Otherwise not_found 'missing: throughput for {gpu}'. Whole-instance billing:
instances = ceil(gpus_per_job / gpus_per_instance), all allocated GPUs billed (9 GPUs on an 8-GPU-only
offer bill 16). Reserved and held-stale rates are never ranked; spot/interruptible only if allow_interruptible.
Each result has rate_id, source_url, verified_at, data_age_hours/degraded/stale. Response has assumptions,
citations, missing_costs, budget_verdict (unresolved: extras unknown) and a quote_id. Compute subtotal only.
```

### `optimize_agent_stack`

- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
Estimate a monthly stack cost from a short description of the agent being built.

description: free text, e.g. 'scrapes 20k JS pages/month via 5 GB residential proxies, 3k search queries,
summarizes with an LLM, budget $300/mo'. Parsed deterministically (keywords/regex, no LLM) into components:
gpu, scraping, proxy, search, llm_api, email, hosting, database, monitoring. Optional fields override the parse:
gpu_hours (GPU-hours/month), gpu_model ('H100', 'A100', 'RTX 4090'), model_size_b (billions of params; sets a
minimum VRAM using FP16 x 1.2), scrape_pages, js_rendering, proxy_gb, search_queries, emails_sent, llm_api,
monthly_budget_usd. Each component gets the cheapest option with a cited price among tracked providers, with
calc, source_url, verified_at, referral_disclosure (set when the link is a referral link) and manual-signup
notes. Components with no cited price say 'no sourced price tracked' (nothing is estimated). Totals show the
math and say whether they are complete. Pulled listings are excluded. Every option has data_age_hours, degraded (> 48 h) and stale (> 7 days); GPU prices older than 168 h never rank by default (listed in the line's stale_prices); max_data_age_hours replaces that default and drops older options (excluded_too_old counts them). Same data as optimize_stack,
compare_gpu_prices and the deal feed.
```

### `optimize_stack`

- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
Cheapest suitable stack among tracked providers for monthly needs (set the quantities you need, rest 0).

scrape_pages: pages/month without JS rendering. scrape_pages_js: pages/month needing JS rendering.
proxy_gb: residential proxy GB/month. emails_sent: transactional emails/month. search_queries: search API
queries/month. gpu_hours: GPU-hours/month (1 GPU), optionally gpu_model e.g. 'H100', 'A100', 'RTX 4090'.
Each line has the chosen plan, monthly_cost_usd, calc, vendor quote, source_url, checked_at and assumptions,
plus alternatives. A vendor price that isn't published is 'unknown' (never estimated). manual_signup_only=true
means a human must sign up (never automate it); exclude_manual_signup=true drops those plans. Same data as
https://thelanternworks.com/api/v1/stack-pricing.json and /tools/stack-optimizer/.
```

### `rebuild_quote`

- The description's first verb is "reopen": a word for a change.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.
- Findings: warn annotation-conflict

Description, as the server wrote it:

```untrusted-server-text
Reopen a past quote. Pass the `receipt` object returned by optimize_stack, optimize_agent_stack or
estimate_job_cost. Returns receipt_intact (unkeyed sha256 matches), reproduced (true only if each line's subtotal
recomputes AND its rate_id and price match the pinned feed revision or that date's published daily snapshot;
"arithmetic_only" with provenance_verified=false when that can't be confirmed), per-line provenance, and
revalidation against today's feed: per GPU line unchanged | price_changed | evidence_changed | held_today |
not_in_feed_today (flat plans: not_revalidated), with revision 'same' or 'new'. Nothing is re-ranked.
```

### `watch_price`

- Nothing in its name or description says what it does, so it is counted as a write.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
Hosted only. Alert your webhook (signed POST) after the daily refresh when a tracked rate for `gpu` crosses at or below max_hourly_usd, or a change of a watched kind (down/up/new; default down) lands at or below it. Only ELIGIBLE rates match (same contract as compare/estimate): every word of gpu must match the model; 'A100 80GB' or vram_gb=80 never matches a 40GB row; tiers (on-demand, community, marketplace, interruptible/spot, dedicated-monthly, reserved; default all but reserved) never matches another class; a row confirmed sold out never matches (in_stock='require' also skips unknown stock). region: code from availability.available_regions; a rate with no unexpired region evidence stays unresolved and never matches. Each event is delivered at most once per watch (dedupe on event_id). Pass api_key or an Authorization: Bearer header.
```

## Only reads (6)

Looks things up and changes nothing. Claude Code: allow.

### `compare_deals`

- The name "compare_deals" says "compare": a word for reading.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
Compare 2-5 deals side by side (price, unit, offer type, code, verification), best value first.
```

### `compare_gpu_prices`

- The name "compare_gpu_prices" says "compare": a word for reading.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.
- Findings: info long-description

Description, as the server wrote it:

```untrusted-server-text
Compare tracked GPU rental prices for one model (e.g. 'H100', 'A100 SXM', 'RTX 4090', 'L40S'), cheapest first. If a result looks wrong, give the user its report_issue_url.

Same data as https://thelanternworks.com/api/v1/gpu-prices.json. Sorted by hourly_usd ascending; each result
has source_url and checked_at. 'Cheapest' means cheapest among tracked providers, not the whole market.
limit 1-25. Unknown model returns not_found listing tracked models.
H100/A100: cheapest_by_form_factor gives one pick per form_factor (tier_class, interruptible, in_stock); same rule as the site. 'NVLink' rows are form 'not stated': NVLink is the bridge, and the vendor doesn't say SXM or PCIe.
Each result has tier_class: on-demand | community | marketplace | interruptible (spot/preemptible) |
reserved | dedicated-monthly. Reserved (reservation/commitment 'starting from') rates are left out unless
include_reserved=true, and even then they are never reported as cheapest. DEFAULT on_demand_only=true: results[0] is the cheapest
on-demand rate. Pass on_demand_only=false to include spot/interruptible, community and marketplace capacity (each
such row is labelled by tier_class and interruptible); the summary then names the cheapest on-demand rate when the
overall cheapest isn't. If a model has no on-demand rate, or every on-demand rate needs more GPUs than requested,
the default falls back to all rates and says so in on_demand_fallback.
gpus (optional, 1-1024; default 1, stated in assumptions): the GPU count you need. A rate ranks only when its
min_gpus (smallest allocation sold) is <= gpus; rates sold only as larger allocations (e.g. 8-GPU nodes) are NOT
ranked and are listed in excluded_min_gpus (vendor, gpu, tier, min_gpus, gpu_count, hourly_usd, rate_id, source_url,
checked_at, reason, min_bill_usd_per_hour, min_bill_basis); min_gpus_unmet counts them (also in brief). Pass gpus=8 to
compare 8-GPU nodes. If every rate needs more GPUs, the error says which gpus value ranks them.
Each result has min_bill_usd_per_hour = min_gpus x hourly_usd (min_bill_basis: hourly_rate, per_second_rate or
monthly_equivalent) and effective_at, when the vendor started charging this price (effective_at_basis vendor_stated,
or first_observed = first daily snapshot showing it after a different price, an upper bound; null when not datable),
separate from last_verified. Neither affects ranking.
Full results also include batch_due_ledger: each scheduled build (gpu-refresh, lium-collect, scout-import,
scout-wider) has expected_at, last_expected_at, and the member_rate_ids it covers, derived only from the
workflow schedules. Brief mode omits the ledger.
A failed or out-of-stock probe keeps failure_reason (sold_out, region, quota, or other) and
failure_reason_detail; both stay on a lapsed claim next to last_known_in_stock. Null when the source
does not say; never guessed. Display only.
A row with unknown stock, and each stale_prices entry, has next_check_at (next scheduled check, UTC), will_backfill
(true: a scheduled check covers it; false: scope limit, nothing will re-check it; null: not derivable),
backfill_eta and next_check_basis (the schedule it came from).
include_history=true adds each result's price_history (daily-snapshot change points) and history.recent_changes
(price changes for this model among tracked providers, last 30 days) from /api/v1/gpu-prices/history.json.
Each result has data_age_hours (hours since its price was last verified), degraded (> 48 h) and stale (> 7 days).
A degraded row is shown and flagged but never ranked first while a fresh row exists; if none is fresh,
stale_fallback=true and the summary says so.
By default a rate whose price is older than 168 h is never ranked or named cheapest: it is listed in stale_prices
('stale price (N days old)'); if every rate is stale, results is empty and the summary says no current cheapest.
max_data_age_hours (optional) replaces the 168 h default and drops older rates; excluded_too_old counts them.
region (optional… [1717 more characters not shown]
```

### `get_deal`

- The name "get_deal" says "get": a word for reading.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data. The description's first sentence says "get" and "url": it reads web pages, which anyone can write.

Description, as the server wrote it:

```untrusted-server-text
Get one deal's public record by id, including redeem_url and agent_instructions.
```

### `get_price_changes`

- The name "get_price_changes" says "get": a word for reading.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
GPU rental price changes among tracked providers, each with old/new hourly price, change_pct, direction,
the vendor source_url, a stable event_id (dedupe key) and rate_id.
Marketplace median moves also have composition_flag (composition, price, mixed, unknown, or null) from
successive median, minimum and listing count; composition, mixed and unknown are not price moves.

since: 'YYYY-MM-DD' returns every change on or after that date, newest first (daily snapshots in
/api/v1/gpu-prices/history.json); omitted returns the latest day's moves (/api/v1/gpu-price-changes.json).
cursor: pass next_cursor from a previous call to get every change after it, oldest first, with has_more;
loop until has_more is false, and store next_cursor only after processing (a watcher never misses or repeats
events). error cursor_expired means re-sync with since and dedupe by event_id.
kinds: optional subset of down, up, new, removed, same. gpu: optional model filter (substring, e.g. 'H100').
limit 1-200. Includes data_checked_at and next_refresh_utc (prices refresh daily at about 10:23 UTC).
```

### `list_watches`

- The name "list_watches" says "list": a word for reading.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data.

Description, as the server wrote it:

```untrusted-server-text
Hosted only. Your watches and per-key alert counts.
```

### `search_deals`

- The name "search_deals" says "search": a word for reading.
- Tag private-data: The description's first sentence says "search" and "email": it reads mail.
- Tag untrusted-input: The server marks it openWorldHint: true: it reaches things outside the server's own data. The description's first sentence says "search" and "email": it reads inbound email, which other people can write.

Description, as the server wrote it:

```untrusted-server-text
Search agent work-stack deals (compute, search, email, unblock, hosting...).

offer_type: verified_discount | listed_discount (not confirmed; verified_at null) | free_tier | trial | standard_referral | credit_only_referral.
max_price excludes deals with unknown price. limit 1-25. Pass next_cursor back as cursor to page.
```

## Claude Code permissions

For `.claude/settings.json`, assuming the server is installed as `lantern-works-deals`:

```json
{
  "permissions": {
    "allow": [
      "mcp__lantern-works-deals__compare_gpu_prices",
      "mcp__lantern-works-deals__get_price_changes",
      "mcp__lantern-works-deals__search_deals",
      "mcp__lantern-works-deals__compare_deals",
      "mcp__lantern-works-deals__get_deal",
      "mcp__lantern-works-deals__list_watches"
    ],
    "ask": [
      "mcp__lantern-works-deals__optimize_stack",
      "mcp__lantern-works-deals__optimize_agent_stack",
      "mcp__lantern-works-deals__estimate_job_cost",
      "mcp__lantern-works-deals__rebuild_quote",
      "mcp__lantern-works-deals__create_watch_key",
      "mcp__lantern-works-deals__watch_price"
    ],
    "deny": [
      "mcp__lantern-works-deals__delete_watch"
    ]
  }
}
```

Capability tags behind the rules (a read tagged outbound is asked about):

- `mcp__lantern-works-deals__compare_gpu_prices`: untrusted-input
- `mcp__lantern-works-deals__get_price_changes`: untrusted-input
- `mcp__lantern-works-deals__optimize_stack`: untrusted-input
- `mcp__lantern-works-deals__optimize_agent_stack`: untrusted-input
- `mcp__lantern-works-deals__estimate_job_cost`: untrusted-input
- `mcp__lantern-works-deals__rebuild_quote`: untrusted-input
- `mcp__lantern-works-deals__search_deals`: private-data, untrusted-input
- `mcp__lantern-works-deals__compare_deals`: untrusted-input
- `mcp__lantern-works-deals__get_deal`: untrusted-input
- `mcp__lantern-works-deals__create_watch_key`: untrusted-input, outbound
- `mcp__lantern-works-deals__watch_price`: untrusted-input
- `mcp__lantern-works-deals__list_watches`: untrusted-input
- `mcp__lantern-works-deals__delete_watch`: untrusted-input

## Remote posture

What an anonymous client saw at the remote endpoint, last 2026-11-20T18:24:02.635Z, unchanged since 2026-11-06T04:21:09.636Z. No credentials were sent and no tool was called. Values are shown as the endpoint returned them. Method: https://radar.anlyon.com/methodology#remote-posture

| Fact | Observed |
| --- | --- |
| Reachable | `yes` |
| HTTP status of an anonymous initialize | `200` |
| tools/list served without credentials | `yes` |
| Tools that change things, listed without credentials | `7` |
| Protocol version | `2025-06-18` |
| WWW-Authenticate on 401 | `not observed` |
| resource_metadata in WWW-Authenticate | `none` |
| RFC 9728 protected resource metadata | `none` |
| Authorization servers | `none` |
| Legacy HTTP+SSE transport | `no` |
| Dynamic client registration | `not observed` |
| Client ID metadata documents | `not observed` |

- **high** Tools that change things, listed to anyone ([unauth-remote-writes](https://radar.anlyon.com/methodology#unauth-remote-writes))

```untrusted-server-text
The endpoint served tools/list without credentials, and the list has 7 tools that change things (1 destructive, 6 write). Whether a call also works without credentials was not tested: tools are never called.

https://lantern-works-mcp.lantern-works.workers.dev/mcp?ref=mcp-registry: tools/list returned 13 tools to an anonymous client, including delete_watch (destructive), optimize_stack (write), optimize_agent_stack (write), estimate_job_cost (write), rebuild_quote (write) and 2 more
```

Posture changes:

- 2026-11-06T04:21:09.636Z: Tools that change things, listed without credentials `6` -> `7`
