What shipped, by version

The workbench reads as one story

Reorganised around the six questions an advisor actually asks, in the order they ask them, with each concept rendered once.

• The page used to be organised by artifact type rather than by question, so the same thing appeared several times over: the seven methods showed up in five places, the six quality factors in five, your peer position in two, and two pairs of sections shared a heading word for word. Now each idea has one home.
• What is it worth? keeps the headline and the fair-value range, and drops the stat rows the chart below already showed.
• Why that number? holds the plain-English explanation, the method-spread chart, and — behind disclosures, because most people never need them — the formulas, weights and confidence for each method, and the blend sliders.
• How strong is the book? now reads as a single argument: the six factors with their scores, what each gap is worth in dollars beside it, and the priced moves to close them directly underneath. The dollar figure used to live only inside a chart.
• How do you compare? keeps the peer chart and folds the old benchmarks panel's extra facts — your implied fee on assets, median staffing — into its footer. The InvestmentNews operating medians sit behind a disclosure, and the sourced-defaults switch moved up to the inputs where it belongs.
• How much rides on the multiple? gathers the sensitivity view and the deal-market context under one heading, and What changed? keeps the comparison to your previous run.

Artifacts that carry the house rules

Nine workflows can now hand you a visual artifact — and the rules that govern the written report govern it too.

• Each of the nine declares an artifact contract, served alongside its instructions: when to offer one, what it is called, what sections and columns it has, and what it must never show. Practice valuation, Peer position, Value driver plan, Succession readiness, Buy-sell funding gap, Acquisition pipeline, Target comparison, Diligence request list, and Deal table read.
• An artifact is offered after the full report and after the What-next menu, never in place of either, and only when the run produced something worth keeping. A re-run with no movement does not get an offer.
• It is layout only. Every figure on it came back from a tool in that session, unchanged — no computing, no re-rounding, no filling a blank cell. Caveats move into the footer rather than being dropped to tidy the design, and positions stay positions: no grades, no ranking, no picking a winner.
• Each one carries its own provenance — engine version, index version, data confidence, the date it was generated, and the line saying it is an estimate for planning discussions rather than a formal appraisal or an offer. An artifact outlives the conversation and gets forwarded, so it has to argue for itself when you are not in the room.
• Some rules are specific to the workflow. The buy-sell gap artifact names the shortfall and stops — it will not recommend a policy, a carrier or an amount of coverage. The peer position artifact is not offered at all when no cohort data came back, and every percentile on it carries its cohort definition, cohort size and filing date. The deal-table read allows no invented quotes: each seat is a lens on the computed run, not a character.
• The look comes from the same colour and type tokens the server uses to render the chart pack, so an artifact and a chart inside it are the same brand by construction rather than by two definitions that drift apart.

Keeping the published multiples current

The weekly check now asks whether a newer edition exists, rather than whether the numbers on one page moved.

• These are dated publications. Re-reading the same post and comparing figures could only produce noise, and it did: a curated row meant the check flagged a change every week that was not one. The check now looks for a newly published edition instead.
• Four publishers are watched rather than one — Sica Fletcher, whose index prices the valuation, plus Echelon, CT Acquisitions and Succession Resource Group, whose figures are shown for comparison and never priced off.
• When a newer edition appears it is fetched, read, and staged beside the edition in use for a side-by-side comparison. Nothing switches on its own; the change is reviewed and then activated or discarded.
• A publisher that refuses automated requests is reported as exactly that, with a date to check by hand, rather than failing quietly. If a page's layout changes so the figures cannot be read, nothing is staged and the change is flagged — better a loud gap than a quiet wrong number.
• On its first real run the watcher found Succession Resource Group's 2026 report, one edition newer than the one on file. Its figures are staged for review; nothing shown on the site has changed until that is activated.

No reconnect, no figure change

• The artifact contract is an added field on responses that already existed, and the skills that carry it are database rows served fresh on every call, so every connected client sees all of it with no reconnect. No tool was added or removed and no tool input changed. Server 1.7.5; the valuation engine stays at 1.3.1 and no computed figure moved.

Four buy-side workflows

For a buyer or consolidator finding and sizing targets — read from the same computed run, never a recommendation to transact.

• Find and screen acquisition targets — give it an acquisition thesis and, with advmcp.ai connected, it drives advMCP's RIA Acquisition Screen for a ranked shortlist of named advisers from filed Form ADV data, then puts a low-confidence screening ceiling on each firm here and hands the survivors to the diligence list. advMCP's RIA Acquisition Screen now points its next step straight into this workflow.
• The seam is stated up front: Form ADV has AUM, client counts, custody and disclosures but no revenue or profit, so every screening ceiling rests on cited-median assumptions and reports low confidence — not a valuation until the target shares its own financials. Nothing in a filing says a firm wants to sell: the shortlist is candidates to research, never a recommendation to approach, offer, or transact, and the ceiling is a walk-away, not an offer.
• Which of these is the better buy? — two or more targets valued under identical settings and laid side by side on price, walk-away ceiling, book quality, concentration, continuity and data confidence, with the tradeoffs named and no winner picked.
• What should I ask for before I buy? — the documents and data a buyer or their counsel commonly request before closing, tailored to which of the target's figures are owner-stated versus system-sourced, and mapped to the valuation figure each one would confirm. A request list, never legal or M&A advice.
• What does this add to my book? — a target read against the buyer's own practice as a combined position: blended AUM tier, how concentration and client mix shift, where the two books overlap. Arithmetic on two computed records, never a projected synergy or a combined price.

Three succession workflows

For an owner planning the handoff — the continuity picture already inside the valuation, with structural questions routed to the owner's attorney, CPA, and insurance advisor.

• What to put on paper before you hand it over — the document-level moves that raise transferability (a named successor, a written plan, a funded buy-sell), each with the engine's own dollar effect, because these are paperwork rather than business change.
• Does your buy-sell still cover the number? — an existing buy-sell's owner-stated funded amount compared to the practice's current value, the gap named in dollars and routed to the attorney and a licensed insurance advisor. It surfaces the gap; it never recommends coverage or designs the agreement.
• What does this number mean for an internal buyout? — the fair estimate and range read for a G2 or partner buy-in: what tends to be financeable versus seller-carried, and the questions that set the terms. It describes what the number implies and never designs the deal.

A catalog you navigate by goal

Twenty-one skills is too many for one flat list, so the catalog now groups by what you're trying to do.

• Every skill carries theme tags — understand, sell, buy, succession, data — returned by both list_skills and get_skill, so an agent can filter the catalog the same way the site does.
• The homepage catalog is grouped into themed tabs with a search box above them; search spans every theme so it never hides a match. The active tab and the search term live in the URL, so a themed view is a shareable link.
• The post-valuation "What next?" menu is grouped by the same themes, recommended workflows first, with the new buy-side and succession workflows suggested only when the run gives a concrete reason — a buy-side lane, an unfunded buy-sell, a near-term retirement.
• Switching a tab or typing in search now holds your place on the page instead of jumping back to the top.

No reconnect, no figure change

• The seven new skills are database rows served fresh on every call, and the theme tag is an added field on responses that already existed, so every connected client sees all of it with no reconnect. Server 1.7.3; the valuation engine stays at 1.3.1 and no computed figure moved.

The % of assets method now uses a published rate — figures moved

Engine 1.3.0. Until now the AUM Percentage method applied a flat 1.5% of assets that was written into the engine, while the method carried the second-highest confidence in the blend. Sica Fletcher publishes AUM percentages by firm size on the same page the index already cites, so the method now reads them instead. Every valuation's % of assets figure changes, and with it the blended fair value.

• The published "Assets Under Management (AUM) Percentages for RIAs" table joins the index as a third axis (keyed on assets, alongside the revenue and EBITDA axes) at index version 2 — the same 2024 edition, same publisher, additional data. Because the engine does not know a practice's RIA type, the five published types become the width of the range: small 0.9–1.3% (midpoint 1.10%), mid 1.4–1.7% (1.55%), large 1.9–2.3% (2.10%).
• Each practice resolves the midpoint for its size tier — small below $225M of assets, mid up to $1.6B, large above — following the same tier-cut convention already used for revenue and EBITDA. Practices under roughly $225M move down, larger practices move up.
• Confidence is now earned rather than assumed: 88 when the rate came from the published index, 60 when the practice falls below the smallest published anchor or no tier applies and the engine constant is used — the same treatment EBOC gets when it runs on a placeholder. In that case the run says so in its own words and lists the rate among the figures no publisher supplies.
• The 1.5% constant survives only as the out-of-tier fallback, renamed so it cannot be mistaken for a sourced figure, and the method note now states which basis it actually used.
• The sensitivity strip sweeps the published AUM range along with revenue and EBITDA, so it reflects the full published spread instead of understating it. Owner-supplied figures stay fixed as before.
• Runs computed under engine 1.2.0 are not comparable to 1.3.0 runs. Valuation Delta already treats a different engine version as non-comparable and offers a matching rerun. No saved run was altered.
• Reference figures for the sample practice: fair value $5.63M to $5.67M, asking price $6.20M to $6.23M.

Multiples: one sourced index prices the number, every other publisher is shown beside it

Everything about multiples in this release, in one place. The engine prices exclusively off the versioned Sica Fletcher SCF Index for the firm's size tier; other publishers' figures are carried for comparison and can never reach a calculation. Owners are never asked for multiples. Engine version stays 1.2.0.

• Revenue and EBITDA multiples default to the index midpoint for the practice's size tier; Comparable Transactions prices off the tier low times the quality multiplier; EBOC keeps a flagged placeholder at reduced weight. The result carries resolvedMultiples and multipleSensitivity, and saved runs stamp the engine version, so Valuation Delta treats an engine change as non-comparable.
• The published multiples are a sourced, versioned dataset rather than a constant. The Sica Fletcher SCF Index ranges (2024 data) and the Echelon Partners deal-activity figures live in versioned tables, and every surface — get_benchmarks, the engine, and the workbench — reads the active version through one loader; the built-in copy is only an offline fallback, and the response says which was used.
• Every Monday 09:00 ET the publisher pages are re-read. An unchanged page only refreshes the verified date; a changed or unreadable page is staged for review in the admin "Index updates" panel, where the owner can diff active vs staged and Activate or Discard. Nothing switches automatically and every check is logged.
• Saved runs record the index version they priced off. Valuation Delta treats a different index version like an engine change — not like-for-like — and offers a matching rerun. Existing runs were stamped index v1.
• The source is shown wherever the number appears: the workbench headline links to the index with its as-of and verified date, the chat headline card names the published index and its verified date rather than implying deal comps, and the sensitivity strip cites the source.
• Owners are not asked for multiples anywhere. Every skill, the compute_valuation report, warnings, engine notes, the get_benchmarks caveat, the server instructions, and the workbench copy were swept of lines that sent an owner to find multiples. The engine supplies them and says so; multiples an owner volunteers are accepted and labeled; any adjustment is a workbench choice with a "Reset to index" affordance, never a task. EBOC keeps its placeholder flag, framed as the system's gap, not the owner's. A test fails the build if the retired phrasings reappear.
• get_benchmarks also returns every other published multiples source that matches a practice, under `crossCheckMultiples` — CT Acquisitions' 2026 RIA M&A ranges tiered by AUM band, and Succession Resource Group's 2025 Advisor M&A Report averages across 176 deals completed in calendar 2024 representing over $13.3B in assets. Display only: shown for comparison in *What's the market doing?*, never an input to a calculation, and the table refuses to be a pricing input at the database level so a future change that tried to wire it into the engine fails there rather than quietly moving everyone's valuation.
• Those cross-check figures are kept in their own versioned table because they are tiered on assets under management rather than on firm size by revenue and profit — mixing the two axes would corrupt both. A source that publishes averages across all firms rather than ranges by size is flagged as not size-matched and labelled "all firms" wherever it appears. Reviewable in the admin panel alongside the index and market-activity datasets; the weekly automated check still covers the index and market activity only.
• *What's the market doing?* moves to 1.2.0 with a "second opinion on the multiples" section: every matching source with its publisher, that publisher's side of the market and its sample; size-matched bands kept visually separate from all-firm averages; what a range would do to this practice's own figure as plain arithmetic; where the pricing index sits among the published sources; and a closing reminder that the number did not change. Six new guardrails: never blend or split the difference between sources, name the axis and size-matching differences first when they disagree, name each publisher's side of the market, never present a cross-check as deal comparables or as equal in standing to the index, never call the index too low or the estimate understated, and never present an all-firm average as what firms this size fetch.
• The homepage source cards now build themselves from the live tables and are grouped by role — what sets the multiples, what positions the practice, and what is carried for comparison only — so a newly added source appears on its own.
• Two standing context sections appear on every first screen: How you compare to firms like yours, and The market you'd be selling into. Workbench and compute_valuation share one compute path and one default (confidence weighting), and a parity test pins both surfaces to the same figures. No valuation figure changes.

Peer cohort folded in

Cohort context arrives with the benchmarks, so an assistant does not need a second server connected to get it.

• get_benchmarks now folds in an SEC Form ADV peer cohort from ADVmcp under advCohort (AUM band, optional state), so cohort context is available without connecting advmcp.ai.
• The cohort is best-effort: if ADVmcp is unreachable or returns a cohort that does not match the requested band, advCohort reports available: false with a reason and every other benchmark field returns unchanged.
• Clients connected before this release are working from a cached tool list and will not see the new `state` input until they reconnect the server.
• The cohort lookup gets 8 seconds rather than 3.5 before it is given up on, and reports how long it took and why it failed, so a slow-but-healthy lookup lands instead of silently reading as unavailable.
• No engine or figure changes; engine version stays 1.2.0.

Three new guided skills

Served fresh from the skills catalog, so they reached every connected client with no reconnect. Each closes with the same plain-English ladder as the existing workflows.

• Reconcile the Numbers — walks a figure that looks wrong back to the field, the source that supplied it, and the method it drives, so a disputed number is settled against the record rather than re-argued.
• Transition Risk Read — reads what happens to the practice if the owner steps away: successor presence, plan and funding, owner-held relationships, and advisor tenure, priced through the longevity factor.
• Market Context — puts the run in the market it would sell into, using the cited index range for the firm's size tier and Echelon Partners' deal-activity context, never as a deal comparable.
• The What-next menu after a valuation now stars a skill conditionally on the run's own signals — plausibility findings, engine-default placeholders, size-tier mismatch, and longevity gaps — so the suggested next step follows the data rather than a fixed order.

Open this run in the workbench

Every valuation result carries a deep link to the same practice and saved run in the web app, so the chat and the workbench are two views of one number.

• compute_valuation returns structuredContent.workbenchUrl (practice and, for saved runs, the run id); list_practices carries the same URL on every history entry. The workbench accepts the link, signs you in if needed, preselects the practice, and loads that run's inputs and weights. Unknown ids fall back to the normal workbench with a notice. Server 1.2.2.
• In the report the link sits directly under the firm name, with the address behind the word workbench, so it is the second line an advisor reads. Server 1.2.4.

Verbatim by construction

Consuming assistants were condensing the report. It is now the tool's only text output, bounded by markers, so there is nothing else to summarize.

• compute_valuation's text is exactly: an instruction line, the report between === BEGIN REPORT === and === END REPORT ===, then the What-next menu. The JSON payload left the text and lives only in structuredContent, which also carries the bare report and the instruction. validate_practice follows the same convention. Server 1.2.3.
• The instruction to reply with the marked text verbatim is now the first line of the server's own instructions (loaded by every client at connection), and the presentation-contract and practice-valuation-run skills say the same. A test runs the real handler and asserts one text item, no JSON, one marker pair, and report-in-text equal to structuredContent.report.
• Integrations that parsed the JSON out of the text need to read structuredContent instead; that is the only breaking change in this group.

Sources and completeness in the readout

The confidence phrase gets its number back, and the report names every system that supplied a figure.

• The headline reads, for example, Built on a complete set of inputs (14 of 14, 100%). A new section, Where the numbers came from, lists each connected source with the plain-language fields it supplied and its last sync date, any owner-stated fields no platform covered, engine-derived fields, provenance conflicts (you said X, the platform reports Y, the report uses which), and the benchmark sources with their as-of dates and cohort size. It closes with the completeness count, the engine version, and the index version. validate_practice renders the identical section, so pre-flight and readout agree. Server 1.2.5.

A layout that reads

Same words, real structure: headings, rules, and one idea per paragraph, rendered correctly by chat clients.

• Firm name as a heading, then the workbench link, then an Asking price heading with the fair estimate and range on one line. Sections are headings, the two data blocks are sub-headings, a rule separates each major part, What this means is four short paragraphs, and every table, list, and rule is surrounded by blank lines. The pre-flight report uses the same rules; a layout test enforces them, and the skills now tell the assistant to preserve headings, blank lines, rules, and tables exactly. Server 1.2.6.

Plain-English ladder with progressive disclosure

Owners and advisors are the primary readers.

• Every workflow skill renders the same ladder: headline card, What this means, What's helping and holding it back with dollar figures, How to move the number, Before you rely on this, and a three-item What next. Each layer ends with Show me the math, which expands the full tables for buyers, bankers, and consultants. Skills use plain titles and a confidence phrase. Server 1.0.11.
• The skills table gained a committed migration and seed file as its source of truth.

The server writes the report

After a rerun collapsed to a one-line result in production, the entire ladder is rendered server-side so no assistant has to reconstruct it.

• compute_valuation returns the complete report in its text and again as structuredContent.report, written in the plain register from the engine's own explainability fields. A rerun is a full report; Since your last run is one line inside it, never a substitute. Raw quality metrics are now persisted on the practice so the fact behind each score reads as a real fact. Server 1.1.2.

Self-healing update surface

MCP clients cache the tool list when the connection is created, so an older connection can silently miss newer tools. The server now makes that detectable.

• `server_info` — reports the server version, the valuation engine version, the live registered tool list and count, the skills catalog size and last update, and a cache note explaining how to refresh a stale connection. Read-only, no authentication needed.
• The tool list in `server_info` is read from the tools the server actually registered at runtime, so it can never drift from a hand-maintained list.
• Server instructions now tell agents to compare their available tools against `server_info` and, on a mismatch, tell the user their client cached an old tool list.
• New skills need no reconnect: `list_skills` and `get_skill` always serve the latest versions from the database.
• `server_info` is callable without authentication, so status pages and monitoring can read the server's live state.
• `benchmark-against-cohort` — places a practice against filed SEC Form ADV peer percentiles while valuing it, with cohort data from advmcp.ai used to sanity-check the figures a valuation was built from and to place the result in context, never to move the number.
• This skill needed no reconnect: skills are served fresh on every call, so a new guided workflow reaches every connected client the moment it is published; only the `server_info` tool required the one-time reconnect.
• Tool count grows from eight to nine; all previous tools are unchanged.

Reliability & trust hardening

The public MCP endpoint and saved valuation records are now more robust, reproducible, and safely framed.

• A fair-use rate limit now protects the public MCP endpoint (per-caller minute, hour and day windows, with a retry-after on 429s). It fails open if the counter store is unreachable, so the service degrades gracefully rather than going dark.
• Every saved valuation now records the deterministic engine version that produced it (engine v1.0.0), so any past run can be reproduced and audited even after the engine's math evolves.
• Valuation summaries now close with an explicit framing line: educational estimate from the deterministic valuation engine — not an appraisal, fairness opinion, or investment, legal, or tax advice.
• Tool errors now return a short generic message instead of raw database detail, with the real error logged server-side.

Reconnect instructions

The in-app reconnect notice is now a two-line nudge that links here; the full reset steps live below.

• Cursor: Settings → MCP, toggle ValuationMCP off and back on (or hit refresh on the server entry).
• Claude: remove the ValuationMCP connector and add it again with the same URL.
• Other clients: disconnect and reconnect, or restart the client so it re-runs tool discovery.
• The MCP endpoint is unchanged, and no new account or key is needed.

New tool: validate_practice

A deterministic pre-flight, run before compute_valuation. Nothing here is model-computed.

• Returns each present field's provenance (system-sourced vs owner-stated), each missing field and its exact point cost toward data confidence, which engine-default multiples are silently in play, any place an owner-stated figure diverges from what a platform actually pulled, currency status against the USD benchmarks, and revenue- and AUM-per-household plausibility against the tier — split into blocking and advisory findings.
• compute_valuation now always attaches the same pre-flight to its result and records an explicit gaps_acknowledged flag, so gaps cannot be skipped — but it never blocks a knowing partial run.
• record_data_source now stores the actual value each platform returned, so a later run can flag when an owner-stated figure diverged from what the platform pulled.
• Tool count grows from nine to ten. This is a tool-surface change, so a client connected earlier must reconnect once to see validate_practice; server_info reports the live count.
• The Practice Valuation Run skill gained a validate step between the benchmark sanity-check and compute.

A house presentation standard, in prose and in structure

Every guided workflow now renders the same way, enforced two ways — a prose contract the assistant reads, and a machine-readable contract a client can render. Layout only; it never changes a computed figure.

• The presentation-contract skill sets three rules: lead with the answer (value, range, confidence together), never buried in a paragraph; put any set of three or more parallel figures — the method spread, quality factors, benchmark positions, data gaps — in a table rather than a comma-run; and collect every caveat into one closing block. Engine-default and cited-median figures are flagged inline in their own cell.
• Alongside the prose rules, every skill now carries a machine-readable output_contract — a format plus the ordered sections its answer should follow (or a style/principle for the reference skills) — returned by both list_skills and get_skill, so a client can render the structure directly.
• Wired into all eight guided workflows and served fresh from the database, so it reached every connected client with no reconnect.

Pre-flight fixes

Two issues in validate_practice, surfaced by real use.

• Quality-factor provenance: the six quality factors now credit the platforms their raw metrics came from. They previously read "owner-stated" even when a platform supplied them, because provenance was keyed on the derived score name rather than the raw metric name recorded against it.
• Data confidence: the server-derived value_per_client no longer counts as a missing field, so a fully-sourced practice reads 100% coverage and clears the "provisional" flag on its own, and a phantom "close data gaps" to-do for a field the server fills in itself is gone. No valuation figure changes — only the reported confidence signal.

New skill: Valuation Council

A completed valuation argued from every seat at the deal table, with a neutral moderator — the seven-advisors council pattern applied to a valuation.

• Eight seats, each a lens on the same computed run rather than a character with invented figures: Seller's Broker, Buyer, Lender, Practice Manager, Succession Planning, Business Partner, a CPA after-tax lens, and The Appraiser's Bar.
• A neutral moderator frames the question the table is convened on, keeps every seat on the computed numbers, names where they agree and diverge, prices the exercise by data confidence, and routes to a credentialed appraiser, CPA, or attorney on a live deal — but never decides or recommends.
• The Appraiser's Bar describes what a credentialed appraiser (CVA, ABV, CBA) would require before signing and when to hire one; it never roleplays the credential or issues an opinion in one's name.
• Served fresh from the database, so it reached every connected client with no reconnect.

Cited transaction-multiple ranges

Transaction multiples were the engine's weakest input — silent defaults with no published source behind them. get_benchmarks now carries cited aggregate ranges.

• get_benchmarks returns a transactionMultiples block: revenue and EBITDA ranges by firm-size tier from Sica Fletcher's SCF Index (2024 data), plus market-activity context from Echelon Partners' 2025 RIA M&A Deal Report — each with source, as-of, and verified date.
• Sica Fletcher publishes tiered multiple ranges (the multiples source); Echelon Partners publishes deal activity, not standardized multiples, so it is cited as market context, never as a median multiple.
• The validate_practice pre-flight now appends the cited size-tier range to the revenue- and EBITDA-multiple "engine default in play" advisories, chosen from the practice's own figures; EBOC and the fixed comparable have no published range and say so.
• Ranges are labeled size-sensitive index figures, not named deal comparables, and are never applied to the valuation — the engine runs on its defaults until you set a multiple from deal data you trust, adopted via upsert_practice and logged as a source.

Guided next step after every valuation

A completed valuation now ends with a menu of what to do next, so the workflow doesn't dead-end at the number.

• `compute_valuation` returns a `nextSteps` menu — the applicable guided skills, each with a one-line "when to use" and a ready-to-send prompt — and renders it as a "What next?" list at the end of the result.
• The menu is context-aware: Data Confidence Triage is starred when confidence is low, Valuation Delta activates only once a second run exists, and Buy-Side Ceiling is flagged as a separate lane.
• Served live from the skills catalog in the home-page order, so new skills and any reordering appear in the menu with no code change and no reconnect.
• The Practice Valuation Run skill now closes by surfacing that menu and running whichever option the user picks — layout and routing only; it never changes a figure or auto-runs an unchosen skill.
• The same menu now appears in the web app — a "What next?" grid in the workbench after a run, and a per-run "What next?" panel on the recorded-valuations page — each with one-click Copy-prompt buttons to paste into your agent.

Guided skills catalog

A public, read-only catalog of step-by-step workflows an agent can discover and follow. No account or key required.

• `list_skills` — browse the catalog of guided skills with titles, blurbs, kinds, versions, and sort order.
• `get_skill` — load a single skill by slug so an agent can follow a written procedure for a valuation task.
• Server instructions now tell agents to check the skills catalog first, so a connected client follows the published procedure instead of improvising one.
• Tool count grows from six to eight; all six v1.0 tools are unchanged.
• 21 guided skills, read live from the catalog:

• Valuation Literacy — How the seven-method blend actually works — what drives each method, what the quality multiplier touches, what data confidence measures, and the vocabulary that keeps a valuation honest.
• Presentation Contract — The house rules for rendering any valuation output — lead with the number, put parallel figures in tables, and collect every caveat at the end. Layout only; it never changes a figure.
• Value my practice — The full sell-side loop: requirements, gathering from the owner's own platforms, provenance, benchmark sanity-checks, and a confidence-weighted blend presented with its caveats in the right order.
• What's missing, and is it worth getting? — Turns a confidence percentage into a to-do list — which missing facts cost the most points, which platform holds each, and what the quality factors are silently assuming in the meantime.
• Do these numbers hold up? — The figures behind a valuation, checked against filed data and published medians — where they disagree, what could explain each gap, and where to go and settle it. Questions for the owner, never corrections to the model.
• What's holding my score back? — Which of the six quality factors is moving the multiplier, by how much, and what the arithmetic says a different score would do — sensitivity as math, never as advice to change the business.
• What happens if you step away? — The continuity picture already inside your valuation, read as one story — who owns the relationships, what is on paper, and how much runway there is — plus an honest account of how little of it this model actually prices.
• What to put on paper before you hand it over — The document-level moves that raise transferability — a named successor, a written plan, a funded buy-sell — each with the engine's own dollar effect, because these are paperwork rather than business change and the cheapest lever an owner has.
• What does this number mean for an internal buyout? — The fair estimate and its range read for a G2 or partner buy-in — what portion tends to be financeable versus seller-carried, and the questions that actually set the terms — describing what the number implies, never designing the deal.
• Does your buy-sell still cover the number? — An existing buy-sell agreement's funded amount, compared to the practice's current value — the gap named in dollars, with the structural questions routed to your attorney and insurance advisor. It surfaces the gap; it does not solve it.
• Find and screen acquisition targets — Turn an acquisition thesis into a ranked shortlist and a screening ceiling for each firm — advMCP finds the candidates from filed Form ADV data, ValuationMCP puts a low-confidence walk-away number on each. A research list, never an approach list.
• How much should I pay? — A buyer-side ceiling estimate: given the practice's economics and the market dataset, what price would leave you with a defensible return?
• Which of these is the better buy? — Two or more practices you might acquire, read side by side — fair estimate, walk-away ceiling, book quality, concentration, continuity and data confidence — with the tradeoffs named and no winner picked.
• What should I ask for before I buy? — The documents and data a buyer or their counsel commonly request before closing on a practice — financials, filings, client agreements, custody, comp, concentration and continuity — as a request list, never a to-do the model priced.
• What does this add to my book? — A target read against the buyer's own practice — the combined AUM tier, how concentration and client mix shift, the blended revenue per client, and where the two books overlap — as the combined position from figures already computed, never a projected synergy.
• What's the market doing? — The published multiples your valuation priced off — which size tier you fall in and why, what the whole range does to your number, which two of the seven methods run on figures nobody publishes, and how current the data is.
• What does this number mean? — Reads a completed valuation the way a seasoned practitioner would — why the methods disagree, which number each side of a deal anchors on, what to probe next, and when the situation has outgrown a modeled estimate.
• How do I move the number? — Where the practice sits against every published value-driver anchor, the arithmetic of each gap, and what the sources say markets reward — positioning and education, with every decision left to the owner and their advisors.
• What changed since last time? — Compare two saved runs and attribute the movement — business figures, data completeness, assumptions, or settings — so a change in modeled value is never mistaken for a change in the market.
• How do I compare to firms like mine? — The first screen already shows your position on three metrics; this skill goes deeper — state and size cohorts, more metrics, the distribution chart.
• Argue it from every side — A completed valuation argued from every seat at the deal table — seller's broker, buyer, lender, practice manager, succession planning, business partner, the CPA's after-tax lens, and the bar a credentialed appraiser would set — with a neutral moderator who surfaces where the seats agree, where they diverge, and where the situation has outgrown a modeled estimate. Each seat is a lens on the same computed number, never a character with invented figures.

Homepage skills surface

The catalog is visible on the marketing site, not just over MCP.

• Live skills section on the homepage, loaded in the route loader from the database so the catalog stays current without a deploy.
• Dedicated icon for each skill, matched to its slug.
• Tools grid showing all eight MCP tools with what each one does.
• A fetch failure degrades to an empty catalog rather than breaking the page.

Reconnect notice

MCP clients cache the tool list at connection time, so a client connected at v1.0 never sees the two new tools.

• Dismissible slide-in notice explaining the caching behavior and the six-to-eight tool change.
• Reset instructions for Cursor, Claude, and other MCP clients.
• The MCP endpoint is unchanged and no new account or key is needed.
• Shows once per browser and links to these release notes.

Data access refinements

• `list_practices` returns full valuation history — label, notes, weights, inputs, quality, methods, quality multiplier, and prior-run link — for the 24 most recent runs when a practice id is supplied.
• Legal entity updated to HnyComb.ai LLC across the Terms of Service and Privacy Policy, with contact routed to hnycomb.ai.

MCP server

The primary interface to ValuationMCP. Agents connect over Streamable HTTP at /mcp and authenticate with OAuth 2.1, so every tool call runs as the signed-in user.

• OAuth 2.1 authorization with dynamic client registration, an in-app consent screen, and per-user scoping on every read and write.
• `list_practices` — enumerate the practices on the account with current data coverage.
• `upsert_practice` — create a practice or merge newly gathered facts into one; omitted fields are left untouched.
• `list_data_requirements` — the full checklist of facts a valuation needs, which method each fact drives, and which kind of platform normally holds it.
• `record_data_source` — log provenance: which platform supplied which field, and when.
• `get_benchmarks` — sourced SEC Form ADV percentiles and published operating medians for sanity-checking gathered figures.
• `compute_valuation` — run the blend and save the result, returning every method's arithmetic, the quality multiplier, the weighted blend, the headline, explainability narrative, optimization recommendations, and prior-run comparison.
• Raw-metric intake: recurring revenue %, organic growth %, retention %, average client age, top-10 revenue share %, and advisor tenure are converted to the six 0–100 quality scores server-side with fixed, published anchors, so identical inputs always score identically across runs and clients.
• Composite quality factors: `quality_longevity` is a succession-readiness score weighted across successor presence, written plan, funded buy-sell, owner relationship share, advisor tenure, and years to retirement; `quality_demographics` blends average client age with the asset-weighted decumulator share.
• Server-side derivation of `value_per_client`: capitalizes revenue per household at the firm's revenue multiple, falling back to the tier median when the firm's own figures are missing, so agents no longer invent a capitalization factor.
• Server instructions that steer agents away from inventing figures — a missing field lowers reported data confidence instead of producing a confident-looking guess.

Valuation engine

Seven methods, each reported with the arithmetic behind it, blended by a normalized weighted mean.

• Revenue Multiple, EBITDA Multiple, EBOC Multiple, Discounted Cash Flow, Client Value, AUM Percentage (1.5% base × quality), and Comparable Transactions (2.2× revenue × quality).
• Six-factor quality multiplier — recurring revenue mix, organic growth, client retention, client demographics, concentration, and succession longevity — clamped to 0.5×–1.5×, with each factor shown at its point of effect.
• User-adjustable method weights plus a "weight by data confidence" preset that loads each method's own confidence score into the weight vector (same normalized-mean math, different weights).
• Per-method data-confidence scoring and a weighted coverage score so a thin dataset produces a visibly low-confidence valuation.
• Sell-side headline: a defensible suggested asking price set above the fair-value blend, with the low/high range across contributing methods.
• Client Value method can use a server-derived value per client when the user does not supply one.
• Manual Workbench for exploring the same math by hand, available to signed-in accounts.
• Deterministic insights engine explains what drove the valuation: revenue scale, quality multiplier, method dispersion, and data caveats.
• Practice-management optimization recommendations are priced by re-running the blend with the improved metric, so each suggestion shows an estimated uplift value.
• Recommendations cover recurring-revenue conversion, succession planning, concentration reduction, client-demographics improvement, and retention programs.
• Automatic comparison to the most recent prior valuation for the same practice: changes in asking price, fair value, quality multiplier, and individual method values.
• Insights are saved with every recorded valuation and surfaced in the workbench, the history page, and the MCP `compute_valuation` response.

Practice workflow

The workspace centers on choosing a firm first, then deciding what to do with it.

• 'Your practices' section so signed-in users can select an existing firm or create a new one on login.
• Inline Rename and Delete actions for practices in both the client workspace and the admin dashboard.
• Cascading delete: removing a practice also removes its recorded valuations and data-source provenance logs.
• Post-selection panel shows the selected firm and offers two clear paths: view the data already on file, or gather new data.
• Inline data viewer displays 'On file' fields formatted as USD or percent, plus 'Still missing' fields with the platform that normally supplies each one.
• Quick links from the selected firm to its valuation history and the manual workbench.
• Practice dropdown at the top of the workbench loads every firm on the account and auto-populates the calculator with its saved inputs and quality scores.

Benchmarks

Only sourced figures — no invented industry data.

• SEC IAPD Form ADV panel derived from the official SEC compilation feed covering 23,512 registered investment adviser firms, with firm-size and client-mix percentiles.
• Published industry operating medians (revenue per client, operating profit margin) used for tier-aware starting points.
• Transaction multiples are explicitly left as user-editable inputs, because neither benchmark source publishes deal multiples.
• Benchmark and industry panels surface the source and attribution next to every figure.

Accounts and workspace

Registration gates the MCP server and the valuation surfaces.

• Email and Google sign-in, with a workspace portal that shows connection status and recorded valuations.
• Connection guide mapping wealth-tech platforms — CRM, portfolio accounting, custodian, billing, accounting, payroll — to the exact fields they supply.
• Supported MCP servers directory with a dedicated page per server covering hosting, authentication, and which valuation fields it can feed.
• Recorded-valuation history with labels, notes, provenance summary, explainability, optimization, prior-run comparison, and delete.

Admin

Output-only oversight across accounts.

• Valuation ledger showing each account that ran a valuation, the provenance of its data, and the calculated results — never the input data.
• Cross-firm averages across every valuation calculated in the app.
• Management of recorded valuations, and a contact-form inbox with reply and delete.
• Client/Admin mode switch inside the logged-in workspace.

Brand, site and privacy

• ValuationMCP identity on valuationmcp.ai: light and dark logo treatments, favicon and touch icon, blue-violet brand palette and gradients drawn from the mark.
• Dark mode by default with a light/dark toggle that persists.
• Landing page positioned as an MCP-native valuation app built for wealth management, with a standardized footer across every page.
• Contact form that routes inquiries without exposing an email address.
• Privacy policy documenting what we store: the practice figures your agent supplies through `upsert_practice`, their provenance, and the calculated valuation outputs. We never copy the underlying records, exports or files from connected platforms, and we never hold your platform credentials. Rights are limited to aggregated, de-identified benchmarks. (Corrected 2026-08-18: this entry previously described raw inputs as never stored; practice figures supplied through `upsert_practice` are stored on the practice record — see the Privacy Policy.)