Skip to content
Liftwire

Remote MCP server and free skill for coding agents

Conversion work your coding agent can run, with the evidence attached.

Liftwire plugs into Claude Code, Cursor, Codex or Windsurf. It audits a live page, plans changes from a library of evidence-graded tactics observed on top SaaS sites, verifies each change once it is live, and measures what happened.

Get a free API keyInstall the skill

Claude Code, one line. Cursor, Codex and Windsurf below.

claude mcp add --transport http liftwire https://cro.nerdcow.co.uk/mcp/http --header "Authorization: Bearer <your-api-key>"
agent session8 seeded tactics in this capture

you › Audit the pricing page at https://cro.nerdcow.co.uk/pricing and plan the three most valuable changes.

▸ audit_page(url: "/pricing")

queued, audit 3ab911f0… poll in 20s

▸ get_audit(audit_id: "3ab911f0…")

ready page_type pricing · LCP 468 ms · CLS 0 · axe critical 0, serious 0 · 1 present, 7 missing

▸ recommend(audit_id: "3ab911f0…", appetite: "days")

ready 3 items. #1 Say that no card is needed [tested] · next to the free plan price or directly under its cta

agent › next it would implement no-credit-card-reassurance in your codebase against its 2 acceptance criteria. Not part of this capture: nothing was changed, so verify ran on a tactic the page already had.

▸ verify_change(url: "/pricing", tactic: "pricing-enterprise-contact-path")

verified_present present before true, after true · 5/5 regression gates passed

▸ get_outcomes(url: "/pricing")

0 outcomes. No analytics source is connected for this site, so changes are recorded as "shipped, unmeasured". Call connect_analytics to connect one.

you ›

Real tool output, captured 2026-10-13 against /pricing on a local stack with a seeded library of 8 pricing tactics. Abbreviated. See the full responses.

How it works

Connect, audit, plan, implement, verify, measure.

Six steps from an API key to a change you can stand behind. The server's own prompts walk the agent through them, so you ask in plain words.

  1. 1

    Connect

    Add the remote MCP server to Claude Code, Cursor, Codex or Windsurf with one API key. The free skill teaches the agent the loop.

    skillMCP

  2. 2

    Audit

    The live page is captured on desktop and mobile, segmented, and compared with the published tactics for its page type. LCP, CLS, axe and CTA position are measured.

    audit_pageaudit_siteget_audit

  3. 3

    Plan

    A ranked plan: each item names a tactic, its evidence grade, the effort, where it goes and the acceptance criteria your agent will test against.

    recommendget_tacticget_examples

  4. 4

    Implement

    Your agent edits your codebase, in your design system, and checks each acceptance criterion. Nothing of yours leaves your repo.

    your agent, in your repo

  5. 5

    Verify

    Once live, the page is captured again: is the tactic there, and did accessibility, LCP, CLS, mobile layout or the CTA count regress? Five gates, pass or fail.

    verify_change

  6. 6

    Measure

    With GA4, Search Console or Clarity connected, each metric is reported for the 28 days before and after. Without, the change stays "shipped, unmeasured". No invented lift.

    connect_analyticsget_outcomessite_memory

The loop in detail·All 11 tools

The product, not a picture of it

What the tools actually return.

Every number below is the committed output of the real tools, run against /pricing on a local stack on 2026-10-13. Where a card shows less than the response, it says so.

get_audit · ready

/pricing audited in 7.1s

pricing
LCP
468 ms
CLS
0
axe critical / serious
0 / 0
CTA above fold
yes
Overflow at 390 px
none
Sections
pricing_table, footer

Present (1)

  • pricing-enterprise-contact-path0.95

Missing (7)

  • no-credit-card-reassuranceTested
  • pricing-annual-toggle-with-savingTested
  • pricing-feature-comparison-tableObserved
  • pricing-highlighted-recommended-planObserved
  • pricing-objection-faqObserved
  • pricing-single-screenObserved
  • pricing-social-proof-adjacentObserved

Captured 2026-10-13; annotation: detectors_only; 8 tactics in the capture library. Abbreviated: ids and usage omitted.

recommend · ready · appetite days

3 ranked changes

  1. #1Say that no card is neededTested

    Where: Next to the free plan price or directly under its CTA. Effort: days.

    • The line sits within the free plan card or under its CTA
    • The statement is true for the free plan
  2. #2Annual toggle with the saving shownTested

    Where: Directly above the plan cards. Effort: days.

    • The toggle switches every price on the page
    • The saving label is visible without scrolling at 1440 px
  3. #3Feature comparison tableObserved

    Where: Below the plan cards (optionally behind a "Compare plans" disclosure). Effort: days.

    • Every plan on the page is a column
    • The table has a caption or heading and real table markup

Captured 2026-10-13. Abbreviated: rationale, structure and copy patterns omitted; the deterministic plan writer ran (no LLM in the capture stack).

verify_change · ready

pricing-enterprise-contact-path

verified_present

present before: true · after: true · detection: detector

Regression gates with before and after values
gatebeforeafterlimit
accessibility pass00critical+serious ≤ 0
lcp pass468264≤ 2500 ms
cls pass00≤ 0.1
mobile_overflow passnonono horizontal overflow at 390 px
cta_count pass77≥ 7

Captured 2026-10-13 on a tactic the page already had, so before and after agree. Before/after images are signed links that expire after 900 seconds (omitted).

get_outcomes · 0 rows

Shipped, unmeasured

No analytics source is connected for this site, so changes are recorded as "shipped, unmeasured". Call connect_analytics to connect one.

Until a GA4, Search Console or Clarity source is connected, a verified change stays exactly that. With one connected, each metric is reported for the 28 days before and after, with session counts, and the record never says more.

Captured 2026-10-13. Verbatim.

The capture used a seeded library of 8 pricing tactics whose grades come from fixture evidence, so the grades shown above illustrate the format, not the evidence for those tactics. The public library is being built and is reported on the sell-readiness page. How the capture was made: scripts/brand/capture-mockups.sh.

Evidence grades

Every tactic says how much is known about it.

Grades are computed from the evidence, never assigned. A number appears next to a tactic only when the grade is measured, and the grade sits next to every claim your agent sees.

Observed

Observed

Top SaaS sites in the corpus use it. This counts adoption; it does not prove an effect.

rule: Seen on corpus sites by two independent annotation passes.

We say
How many corpus sites use it, where on the page, with examples you can look at.
We never say
That it will change your numbers. Popular is not the same as proven.
Tested

Tested

Outside evidence supports it: a published experiment, or tracked sites that added it and kept it.

rule: At least one experiment-grade evidence link, or a change kept on 3 or more tracked sites.

We say
That evidence exists, with links to the sources so your agent can read them.
We never say
A number for your site. Experiments ran elsewhere, on other traffic.
Measured

Measured

The effect was measured on connected sites after the change shipped.

rule: Outcome rows on at least 2 sites, each with 500 or more sessions per period.

We say
The measured change, with the metric, the number of sites and the periods behind it.
We never say
That the same change will repeat on your site. It is a measurement, not a promise.

How grades are earned and lost·Published tactics

Connect your agent

One endpoint, one key, 11 tools.

Paste the snippet for your agent, replace the placeholder with the key from your email, and ask it to audit a page. The skill is optional and free: it teaches the agent the order of the loop, how to read grades, and what it must never claim.

MCP endpoint
https://cro.nerdcow.co.uk/mcp/http
Claude Code plugin (skill + server)
claude plugin install liftwire@liftwire

Getting started·Skill and plugin·Connect page

claude mcp add --transport http liftwire https://cro.nerdcow.co.uk/mcp/http --header "Authorization: Bearer <your-api-key>"

Access

Start free. Pay for sites and audits, not seats.

The free plan covers 1 site and 3 page audits a month, enough to run the loop on a pricing page end to end.

Free

One site, three audits a month

Free

  • 1 site
  • 3 page audits per month
  • Full tactic library
Get a free key

Pro

For a team shipping on its own sites

£39/month

  • 3 sites
  • 30 page audits per month
  • Metered overage
Email us for Pro

Agency

For agencies working across client sites

£149/month

  • 15 sites
  • 150 page audits per month
  • Metered overage
Email us for Agency

Paid plans are set up by hand while online billing is being finished; Pro and Agency overage is billed per audit beyond the monthly allowance. What counts as usage.

Get a free API key

One free key per email address. We send it once, with the setup steps, and store only a hash of it.

Security and conduct

Built to be pointed at sites you run, and polite to the ones you do not.

  • API keys are the only credential

    Keys start with lw_live_, are emailed once and stored as a hash with the prefix and last four characters. No end-user login, no dashboard, no OAuth to grant. Rotation gives the old key a day of grace.

  • Signed, expiring image links

    Every screenshot link an agent receives is signed and expires after 15 minutes. Raw storage keys never appear in a response.

  • Audits stay on the public web

    Public http and https pages only: the worker refuses localhost, private and internal addresses, redirects to them and loopback, and caps the transfer size.

  • Robots honoured, takedown available

    Corpus captures respect robots rules and identify the crawler. The takedown form flags a site at once: the library stops capturing it and its examples are hidden from the public pages and from agents straight away.

  • Rate limits and size caps

    Requests are rate-limited per key with HTTP 429 and a Retry-After header; request bodies are capped at 64 KB; summaries are capped and say what was omitted.

  • Nothing is rounded up

    The corpus thresholds and where the library stands against them are published, with figures reported as not measured until there is enough data behind them.

Limits·Sell-readiness thresholds·Takedown

Questions

The short answers.

Which agents does it work with?

Any agent that can call a remote MCP server over streamable HTTP with a bearer token: Claude Code, Cursor, Codex and Windsurf have setup snippets on the connect page. The skill is packaged as a Claude Code plugin, with rules files for Cursor and Codex.

Does it change my site?

No. Liftwire audits, plans, verifies and measures. Your agent makes the change in your codebase, in your design system, where you review it like any other change.

What counts as usage?

Page audits started by audit_page and audit_site count against the monthly allowance; every audit response carries a usage block. Reads such as get_audit, recommend, get_tactic, get_examples, site_memory and get_outcomes do not count. Starting the same audit again within 24 hours reuses the first result.

What will it never claim?

A number for a tactic that is not graded measured, a promise for your site, or that a change worked before verify_change has checked the live page. The evidence grade sits next to every item, and the skill repeats the rule to the agent.

Is there a dashboard or a gallery?

No. Humans get this page, the docs, an email with the key and, later, billing. Everything else is for agents: the MCP tools, the skill, and one machine-readable page per published tactic.

What happens to the pages you capture?

Captures of your pages are stored for your workspace so verify_change has a before state. Corpus captures of public SaaS sites respect robots rules, identify the crawler, and link to a takedown form that stops further capture of a site.

Something else? [email protected]