Changelog

User-visible changes to the hosted Pro Edition and this site, dated by production deploy — newest first. Internal refactors and dependency housekeeping are omitted. The Free Edition’s npm packages keep their own release notes: elicitly and @elicitly/tools.

2026-10-13

• A checkpoint you control. The Approval Flow guide now explains why approval requests always go to you, and where an enforced sign-off (separation of duties, audit) belongs instead.
• elicit_proposal names the right field. Its description now says a submitted proposal comes back from elicit_result as formSubmission, the field it actually returns, so agents read the included values correctly.
• Trademark Policy. How you may use the Elicitly name and logo, linked from the site footer.

2026-10-10

• elicit_display has worked examples, rendered live on the page: a carousel of Jupiter’s moons with images and Wikipedia links, and a top-salespeople list with photos that falls back to initials. A new section shows what happens in hosts that can’t render MCP Apps: what the model is told, and the Markdown answer it gives instead.
• Diagrams in the docs. New diagrams on the elicit_display page, the Pro overview, the narrated gate guide, and Data Retention. Every diagram now has a text description for screen readers, and opens full-screen when you click it or press Enter.
• A refreshed home page on rich presentation for Agent Skills: a before-and-after of a text reply versus a rendered display, and example workflows for sales approvals, supplier master data, and releases. The Support Matrix chips now show MCP Apps support.
• Pricing covers rich displays. elicit_display is listed for every edition, along with how each edition loads your own images, styles, and fonts (Pro: an organization allowlist through the resource proxy). A new “What Pro adds” section, and elicit_selection is now described as rich selection: templated rows, multi-select, up to 500 options.
• elicit_display renders shrink to fit. A display now gets shorter when its content does (for example after you widen the chat panel), instead of keeping empty space at the bottom. The Free Edition gets the same fix in elicitly 0.8.3.
• Pro organizations can allow https://www.elicitly.ai in their elicit_display allowlist, so the salespeople example works as written. Its sample photos now use that address.

2026-10-09

• elicit_display templates can load images, stylesheets, and fonts from your own origins. An organization admin lists allowed origins (up to 20, wildcard subdomains supported) under Organization in the dashboard. Templates keep their normal URLs: Elicitly serves those resources through its resource proxy, using links tied to your organization. Removing an origin stops it loading within about a minute, past renders included. When a template uses an origin that isn’t allowed, the model is told which origins are, and that an admin can add more.

2026-10-08

• New: show the user rendered content with elicit_display. Pass a template and JSON context, and hosts that support MCP Apps render it inline in the conversation — verified in Claude, ChatGPT, and Claude Desktop. Use it for tables, reports, and summaries the user should scan instead of Markdown. Templates can load Tailwind and DaisyUI from jsDelivr; the render follows the host’s light/dark theme (even when it switches while the render is on screen), offers fullscreen, and opens links in the browser. The rendered HTML never enters the model’s context: the model gets your summary, plus a note when the host doesn’t support MCP Apps or the template references images or scripts from origins the view blocks. The tool is also in the Free Edition (elicitly 0.8.0 and 0.8.1 — release notes).
• Templates have a new home. The Templates guide moved out of the Pro section, since templates now drive both inline display and the Pro review pages. The old address redirects.
• elicit_display where it can’t render. In clients that don’t show MCP Apps, such as the Claude Code CLI, the model is now told the render wasn’t shown and presents the information in its reply instead. The result’s structuredContent carries appsAdvertised and the same notes as the text, for clients that read only structured results. Claude Code in the Claude desktop app renders it inline.
• Tool results can gain fields without breaking connected clients. Every tool’s output schema now allows additional properties, so new result fields don’t fail validation in clients that cached an older tool list. If a call ever fails with “Structured content does not match the tool’s output schema”, reconnect the Elicitly connector to refresh its tools.
• MCP Apps in the Support Matrix. The Support Matrix has a new MCP Apps column showing which hosts advertise MCP Apps (read from each row’s archived handshake) and which have actually rendered elicit_display: Claude, Claude Desktop, and ChatGPT Desktop do; Goose Desktop is marked broken because it drops the tool result’s _meta before the view sees it, now on the tracked upstream issues. The Claude.ai row is refreshed: its connectors now advertise MCP Apps but no longer advertise elicitation. The CSV export includes the new column.

2026-10-03

• Breaking: one result shape for every review tool. elicit_result now returns { status, decision, context, formSubmission } for approvals, selections, and proposals alike, where formSubmission is what the review form submitted. Two fields that repeated it are gone: read a selection from formSubmission.selectedIds (no more top-level selectedIds), and a proposal’s included fields from formSubmission itself (no more top-level values). Each tool’s page now has a Result section showing its submitted and declined outcomes.
• A dashboard built around your requests. The dashboard is wider, leads with your activity, and tucks the agent-connection instructions into a collapsible panel once you have requests. Each row shows the request’s type and a colored status badge, and you can filter by Pending, Submitted, Declined, Canceled, or Expired. Organization admins switch between Mine and Organization activity, with a column showing who made each request. The app header now carries the Elicitly logo.
• Drill into any request. Click a request to see its timeline (requested, opened, decided, collected by the agent), the request your agent sent, and the exact result it received. A selection shows every option with its details and marks the one chosen; a proposal shows each field’s source, current, proposed, and submitted values. Admins can open any request in their organization.
• Bigger templates. Templates can now be up to 32 KB each, up from 20 KB — room for inline SVG icons and small images. The elicit_selection page shows how to reuse one SVG icon across every row.

2026-10-02

• A proposal grid that fits the page. The elicit_proposal review grid could grow wider than the page — a multi-select field laid every option out in one row — pushing the Include checkboxes out of view. The Source, Current, and Proposed columns now share the width equally, long URLs and IDs wrap, and the grid only scrolls sideways on genuinely narrow screens.
• Multi-select fields are checkbox lists. On both the approval form and the proposal grid, a multi-select field now shows one checkbox per option, one per line — no more Ctrl/Cmd-click to pick several.
• Long text gets room to read. A text field whose value is long or spans several paragraphs now shows as a multi-line box that grows to fit its text, even when the request didn’t set a maxLength. Read-only Source and Current cells keep their paragraph breaks too, so they read the same as the Proposed value. See Form controls for how each field type renders.
• Restore Proposed restores everything. After Reject All, Restore Proposed now puts each row’s Include checkbox back as well as its value, returning the grid to how it first opened.
• Complete highlight outlines. The outline marking a Source value that differs from Current is no longer cut off on the last row.

2026-09-29

• A clearer confirmation after you choose. Picking an option in an elicit_selection table used to leave a bare “Selection recorded” line where the option table had been, with nothing to say whether you were finished. The confirmation now reads as a proper result — a heading, a check mark that draws itself in, and the same “You can safely close this tab” reassurance every other review outcome already gave you — and screen readers announce it instead of it appearing silently. Choosing “None of these” to refine a search reads the same way. A request that turns out to have already been decided, or to have expired, gets the clearer wording but no check mark — it would be claiming you’d just done something you hadn’t. If your system asks for reduced motion, the mark appears without animating.

2026-09-20

• Sign in with an emailed code. You can now sign in at app.elicitly.ai with a six-digit code sent to your email address, instead of a password — handy when your company blocks third-party sign-in, or when you’re approving a request from your phone. Existing password sign-in is unchanged, and the code takes you straight back to the review request you opened.

2026-09-19

• Links in review-page templates now open. A target="_blank" link in an elicit_selection row — or in an approval or proposal displayTemplate — was silently dropped by the review page’s sandbox: clicking it did nothing at all, with no error. Links now open in a new tab, and the opened page loads as a normal page rather than in a stripped-down sandboxed one. Nothing else about the sandbox changed — template markup still can’t run scripts in the selection table or reach your session. The Templates page now documents links and images, including why a link without target="_blank" replaces the option table and strands the reviewer, and why images need no host allowlist.
• Selection tables now use the full page width. A multi-column option table was clipped by the narrow reading column, pushing columns off the right edge; the selection review page now uses the same wide layout as the proposal diff grid.

2026-09-14

• Goose joins the Support Matrix — both surfaces, all green. Goose CLI and Goose Desktop (1.50.0) verified live against the hosted server: form elicitation working over remote streamable-HTTP with fast accepts, and MCP 2026-07-28 negotiated natively — the matrix’s first CLI row with a non-null Protocol cell. URL mode isn’t advertised, but the hosted review-page fallback round-tripped cleanly on both. Verbatim fingerprints archived, as always.
• The review page’s post-submit confirmation now respects dark mode. After submitting a selection, the “Selection recorded” panel rendered as a white card with dark text for dark-theme users; it now follows your color scheme like the rest of the page.
• The contribute-fingerprint prompt captures host versions more reliably. If you don’t know your host app’s version, the agent can now offer the embedded MCP client’s self-reported version as a candidate — applied only when you confirm it matches, since wrapper apps (like Claude Desktop) embed a differently-versioned client.
• security.txt now declares UTF-8 explicitly (as RFC 9116 requires), so its punctuation renders correctly in every browser — the llms.txt family got the same treatment.
• A published vulnerability disclosure policy. Security researchers now have a clear front door: how to report, what to expect from us (acknowledgement within 5 business days; public credit for valid reports — no paid bounty program), what’s in and out of scope, and a safe harbor for good-faith research. The policy is referenced from our security.txt per RFC 9116 and linked in the site footer.

2026-09-13

• A proper hello (and a check-in) when you join. New organization admins now get a short welcome email at signup with the fastest path to a first approval, and a personal check-in about two business days later asking how it’s going. Both are plain text from a real person — replies go straight to a monitored inbox. The privacy policy now lists Resend, the service that delivers them.
• The dashboard’s “Connect your agent” section works for every MCP host. Your MCP endpoint URL is now front and center with a copy button, the Claude Code one-liner is labeled as the example it is, and a link leads to the install guide for everything else.
• Install guide: ChatGPT and Codex instructions. A new section points at OpenAI’s MCP connector guide (ChatGPT desktop, Codex CLI, and IDE — one shared configuration), notes the Developer-mode requirement, and links our troubleshooting page for the known Codex OAuth-refresh issue.

2026-09-09

• The site reads better on phones. The header and footer now stack cleanly in portrait, link tap targets got breathing room, and pages use more of the screen width on small displays.
• llms.txt is now linked from the site footer. The machine-readable docs index has always been served (and advertised to crawlers); now it’s one click away for humans too.

2026-09-08

• Clearer Contact us page. Sales and support are now two distinct paths: talk to sales about the Enterprise Edition (self-hosting, SSO/SAML, compliance), or get support with direct links to the documentation and live service status.

2026-09-07

• elicit_doctor no longer fails on hosts that don’t support elicitation. Running elicit_doctor with probeElicitation: true against a host that doesn’t advertise elicitation now returns a clean “not attempted” result (and still records/shares the report) instead of erroring the call — so the capability report always comes back.
• The selection review page matches Elicitly’s design. The hosted elicit_selection grid now uses branded controls, clearer rows, and a primary Submit / quiet Decline layout, in both light and dark themes.
• Clearer elicit_doctor description for better agent tool selection — mirrors the same tightening in @elicitly/tools.
• New Troubleshooting page. Fixes for common connection problems, starting with ChatGPT/Codex connectors that stop working after a failed OAuth token refresh (codex mcp logout + reconnect).
• Elicitation Support Matrix refreshed. Current-build fingerprints for Claude Desktop 1.46388.4 (local and remote connector — elicitation not advertised on either) and ChatGPT Desktop 26.901.51231 (stdio and remote HTTP, both form and URL modes verified working), plus a new attachment-path dimension distinguishing config-file from desktop-extension installs.

2026-09-04

• Richer MCP tool contracts across all 12 hosted tools. Every input parameter now carries its own description (elicit_selection’s Liquid templates and selection limits included), and each tool declares a human-readable title, behavior annotations (read-only / destructive / idempotent hints), and — except the plain-text elicit_ping — an output schema, with results returning structuredContent alongside the JSON text. Agents get better tool selection and typed results; both transport lanes serve the identical contract. elicit_confirm / elicit_form / elicit_doctor mirror the same contract shipped in @elicitly/tools 0.6.0, so hosted and Free stay in lockstep.
• New concepts page: The Narrated Gate. Why a prose confirmation gate can be satisfied by narration — the agent describing asking instead of asking — and how a tool-call gate like elicit_confirm closes that gap.
• Landing page: the hero’s elicitation dialog now carries a slow aura ring spotlighting the human-checkpoint moment (respects reduced-motion).
• Connector icon fix for dark-mode hosts: the site now serves a full-bleed apple-touch-icon.png, so hosts that scavenge favicons for connector chips get maskable artwork instead of the rounded browser-tab favicon whose transparent corners rendered white on dark themes.

2026-08-23

• Claude.ai (Cowork, web) joins the Support Matrix. Like Claude Desktop, it advertises elicitation.form but never renders the dialog (probe advertised_but_unanswered, embedded client claude-code 2.1.241) — the same #88901 hang. This corrects the earlier note that claude.ai lacked elicitation entirely: it advertises it, it just doesn’t work yet.
• Claude Code stdio row refreshed to 2.1.231. A community fingerprint (elicitly 0.5.0, form probe accepted in 7.8s) confirms the row holds on the newer release; the 2.1.228 evidence moves to superseded. Contributed via the free edition’s contribute-fingerprint prompt.
• Claude Desktop joins the Support Matrix — its thesis row. Custom connectors advertise elicitation.form, but the live probe never renders a dialog: verdict advertised_but_unanswered, archived verbatim from an owner-run capture on Claude for Mac 1.34493.1. The embedded client identifies as claude-code 2.1.241 — the same client library that renders forms correctly in Claude Code — so the gap is the Desktop connector UI, not the protocol client. This is the case the Pro Edition’s hosted review page exists for.

2026-08-22

• The connector icon renders cleanly in MCP hosts. Hosts mask connector icons with their own corner radius; the site favicon’s baked rounded-rect showed white corners in Claude Desktop’s connectors list. The MCP endpoint now serves a full-bleed square mark (/favicon.svg, /icon.svg) and declares it via the spec’s serverInfo.icons, so both icon-aware hosts and favicon scavengers get maskable artwork. Browser tabs keep the rounded favicon.
• contribute-fingerprint survives hosts that never answer the probe. Field-testing in Claude for Mac showed two failure shapes the prompt didn’t handle: the prompt text arrives as an attachment (which a cautious model reads as injected instructions), and the live probe can hit the host’s tool-call timeout before the report returns. The prompt now states its user-initiated provenance up front and, on a probe timeout or repeated error, contributes a passive report with the failure recorded in notes — the timeout is the finding (advertised_but_unanswered), not a reason to give up. The prompt’s docs gain a Claude Desktop tip: invoke it from the + menu (Connectors → Add from your connector), and use Cowork mode — Chat mode’s tool calls errored out in testing. And the root cause is fixed on the hosted server: the doctor’s live probe now gives up after 40 seconds — below common host tool-call limits — so a host that never delivers the dialog gets the finished report back in one round, timeout verdict included.
• Pro install docs cover Claude Desktop and claude.ai. The Install page’s register step was written for Claude Code and config-file hosts; it now also walks the Add custom connector form (name + URL, keep the detected defaults, no request headers) and the Connect sign-in step.

2026-08-19

• Contribute a fingerprint in one step. elicit_doctor gains an opt-in share: true (plus productName / productVersion / notes): the hosted server stores the finished report in the Support Matrix’s evidence queue — no copy-paste, no email. The new contribute-fingerprint prompt (the server’s first MCP prompt; hosts that support prompts list it — in Claude Code it’s a slash command) walks the whole flow: host details, consent, probe, share. Reports remain payload-free (handshake echo, capability booleans, probe verdicts); nothing is stored without share: true; the matrix stays curated — contributions are evidence for review, not automatic edits.
• The free edition contributes too — elicitly 0.5.0 ships its own contribute-fingerprint prompt: same flow, but ending in a pre-filled GitHub issue (nothing sent automatically). Its first community contribution landed within hours and gave the Claude Code stdio row its first verbatim archive — every matrix row now links archived evidence.
• Docs: a Prompts section, sibling to Tools, starting with contribute-fingerprint — flow, consent scope, and the planned Free-edition variant. The Support Matrix also now sorts freshest capture first and separates each row’s current evidence from superseded fingerprints (earlier versions of the same host), both on the page and in the CSV.
• Claude Code remote-HTTP form elicitation: working as of 2.1.228. The very first prompt-driven contribution flipped the matrix’s most-watched cell: the auto-cancel bug (anthropics/claude-code#85442) is fixed in practice — live probe, dialog rendered, answer round-tripped. History preserved in the row; before/after fingerprints archived.

2026-08-16

• Listed on the Official MCP Registry and Smithery. Both editions are published on registry.modelcontextprotocol.io under the DNS-verified ai.elicitly namespace — ai.elicitly/elicitly (the free stdio server on npm) and ai.elicitly/pro (the hosted remote). Elicitly Pro is also live on Smithery as elicitly/pro, domain-verified, and the install guide notes the Smithery gateway as an alternate connection route.
• Agent discovery, end to end. An agent can now find, understand, and authenticate to Elicitly starting from nothing but the domain name, by standards at every hop: a DNS-AID record (_mcp._agents.elicitly.ai, DNSSEC-signed — the whole zone now validates) points at the MCP host; an API catalog (RFC 9727) and homepage Link headers (RFC 8288) advertise the machine-readable surfaces; an MCP Server Card (SEP-1649) at /.well-known/mcp/server-cards.json — on this site and on the MCP host itself — describes the server, its endpoint, and how to authorize.
• Content Signals in robots.txt. Explicit search=yes, ai-input=yes, ai-train=yes: for an agent-native product, being searchable, quotable in AI answers, and present in training data is distribution, not leakage.

2026-08-15

• Fixed: form-mode elicitation answers through the hosted server. A runtime restriction in our hosting environment broke validation of every form answer a host delivered over the legacy protocol — the dialog worked, but submitting it returned an error. Found via a ChatGPT Desktop fingerprint, fixed the same day, and re-verified live. (The failure was ours, not the host’s — the correction is on record.)
• elicit_doctor failures now say why. Error probes include the underlying rejection in probes.elicitationForm.reason — the field that cracked the bug above.
• Support Matrix: evidence layer. Rows now link the verbatim elicit_doctor fingerprints they were derived from, split by transport (stdio vs. remote HTTP), track the negotiated MCP protocol revision and capture date, and export as CSV. ChatGPT Desktop is the first host verified working over remote HTTP in both elicitation modes — form and URL.