RXed AI News

The protocol work is the real story — a 37-tool MCP server and an MCP client, both directions, permission-scoped — wrapped around a €4,600-a-seat dialer whose core trick still puts two seconds of dead air in front of every prospect who answers.

Nobody records the kitchen-table conversation better, and the grounding is genuinely excellent — but at roughly €3,700-4,600 per seat with a five-seat floor, no public API and no MCP, you rent the best dataset in field sales and can never take it home.

The most complete healthcare AI stack you can buy as an API, and the only clinical vendor that speaks both MCP and A2A properly. Check the eval and fine-tuning gaps before you build your product on it.

The only ambient voice technology certified UKCA Class IIa, with published safety science to back it. Narrow by design: no MCP, no A2A, desktop-only, and one independent NHS pilot found 20% of licensed users never logged in once.

The only ambient scribe with a real developer platform behind it, and the only one that published research saying the industry's own quality yardstick is broken. Independently validated at $1,223 incremental revenue per provider per month, and honest enough to report a 31% hallucination rate in its own peer-reviewed study.

The one ambient scribe a solo clinician can actually buy this afternoon: posted prices from $39, a 7-day trial, 26,000 clinicians and 32.6 million visits transcribed in 2025. It scores low here because it deliberately ships none of the platform layer, and because nobody outside Freed has ever peer-reviewed its accuracy.

The only eval platform we have scored that also ships a serious open-source gateway — Bifrost, Apache-2.0, 7,300 GitHub stars, 1,000+ models and an MCP gateway — wrapped around per-seat billing and three-day log retention on the free tier, so the bill tracks your headcount rather than your traffic.

The lowest-friction way into voice-agent QA — $0.25 a test minute, no sales call, an MCP server and Claude Skills on every plan — but it grades transcripts rather than audio, its testing is state-blind, and most of the public evidence for how good its evaluations are was written by Cekura.

The fastest self-hosted eval platform to stand up — one pip install, one process, Postgres when you need it — and the best RAG evaluation in the category, but the server is Elastic License 2.0, not open source, and "fully open source, no feature gates" is the vendor's own overstatement.

Mintlify bought it on 03/03/2026 and put it in maintenance mode, and the repository shows exactly that: the last GitHub release was August 2025 and the most recent commit removes an incident banner. Do not start here in 2026, however good the two-line integration is.

The broadest agent platform we have scored, and the one most likely to surprise you on the invoice: Engine rescans every 6 hours at 10-15 LCU a run, which is 60 to 90 dollars a day per project if you leave the spend limit blank.

Apache-2.0 with the whole product in the free self-host and the cheapest managed tier in the category at 19 dollars, and the only platform here that will block a PII leak before it reaches a user, as long as you self-host, because the guardrails run nowhere else.

14,000 customers, 300+ patents and the only AI governance vendor here that already had the enterprise installed base before the category existed. Runtime guardrail enforcement shipped in March 2026 and that is the real move. Read the release notes though: AI Policy Manager and guardrail enforcement were still public preview in the May 2026 release, and the median contract of $11,835 tells you nothing about what an AI governance deployment costs.

The shortest path from an OAuth account to an agent-callable tool: 10,000+ tools across 3,000+ APIs behind one static OAuth MCP URL, with read/write/destructive annotations, real Node/Python/Go/Bash in every step and version history with rollback. Workday now owns it, and a free tier of 100 credits a month is not the customer they bought.

The deepest regulation-to-control translation in the category, and the sharpest thinking about where agent governance actually belongs. But the product that acts on that thinking, Agent Governor, is a research preview on one harness with no production SLA, and the open-source assessment framework Credo AI built its early reputation on has been archived as unmaintained.

The only GRC platform I have audited that is itself ISO 42001 certified, processes AI personal data in Frankfurt under its own DPA, and built its MCP so an agent structurally cannot delete your compliance records. Nine named agents do real work. The hole in the middle is that after three years Scrut still will not name the models running any of it.

The best protocol surface in GRC and the most honest model disclosure I have found in the category: 100+ MCP tools across 38 categories with full read and write, and a support page that names OpenAI GPT-5 generation and says plainly that Secureframe trains no models of its own. The AI itself is thin behind that. There is no eval surface beyond a thumbs-up button, no agent layer, and the hosted MCP runs in the US and UK only.

The most transparent AI stack we have scored in this category, with every model provider named, three vendor guardrail layers switched on and a working MCP server that writes as well as reads. The catch is that the headline agent product was announced in April and still is not shipped, the AI credit meter went live in January, and customer support scores 1.9 out of 10 on TrustRadius against 8.0 for ease of use.

The best MCP implementation we have scored on any SMB tool — read and write, OAuth, org dashboard — bolted onto a notetaker whose advertised $19 seat becomes $77 the moment you want the features you came for.

The best grounded-answer engine we have scored — and you cannot use it. Free to 915,000 verified U.S. clinicians, blocked in the EU and UK by OpenEvidence's own choice, with no API, no SDK and no MCP for anyone.

The best AI guardrail console we have scored in this category: admins set culture, pedagogy and compliance prompts at group level, and a course that contradicts the compliance prompt is stopped before it generates. The AI behind it is text-to-text only by the vendor's own admission, runs one Azure OpenAI model family with no model choice, and MCP is a Q3 2026 promise rather than a URL you can point Claude at.

The cadence engine is still the best in the category and the MCP server is real, but you pay $125-165 per user per month for AI that independent reviewers call helper-grade, you cannot see which model writes your emails, and this is the vendor whose GitHub account was open from March to June 2025 and ended up costing hundreds of Salesforce tenants their data.

The most transparent clinical AI vendor on the market and the only one that publishes a EUR price — but it is walking away from the small practices that made it.

The only ambient scribe that beat a control group in a randomised trial, and the only one with a real public API — but it stopped publishing its price.

The deepest voice-agent testing loop we have scored on tool-call contracts, DTMF and IVR behaviour, and turning a production failure back into a regression test, sold by a company that will not publish a single price and grades other people's agents on a judge model it refuses to name.

Renamed Superhuman Docs on 08/07/2026, this is still the best database-backed doc on the market and now has the best-built MCP server of any workspace tool we have scored, but its AI runs on models the vendor will not name, cannot read a PDF, has no quality measurement of any kind, and its enterprise-search product was wound down into a sibling you have to buy separately.

The deepest eval-to-CI loop in the category, and the only platform here that bills every byte of your traces with no spending cap, so set billing alerts before the first production trace.

The open-source default for LLM tracing, MIT-licensed with 79 MCP tools out of the box, and you pay for it in operations: v3 means Postgres plus ClickHouse plus Redis plus S3, and ClickHouse now owns the company.

The most feature-dense AI layer among work-management platforms — real memory, real multi-model choice, real agents — but every bit of it rides on a mandatory per-seat Brain add-on, and ClickUp's headline agent benchmark is self-graded.

A genuinely rebuilt agent platform bolted onto a mature work-management core — strong governance and MCP support, but the AI still runs on borrowed frontier models and monday's own Q1 2026 earnings call described real agent/MCP usage as 'still not very significant numbers.'

Cognigy (NiCE Cognigy)

Enterprise-grade agentic CX that runs live at Lufthansa and Mercedes-Benz scale — but you'll pay six figures a year for it, and since September 2025 NICE owns the roadmap, not Cognigy.

Dust

The fastest way to wire agents into Slack, Drive and Notion for €29 a seat — until the 1GB/user data cap and Enterprise-only governance push you into a custom quote.

GitHub Copilot

The widest model catalogue and the deepest agent surface in any coding tool — but since 1 June 2026 the $10 sticker price is only the entry fee on a token meter, and heavy agent users are burning a month of credits in a week.

Coval

The most complete voice-agent evaluation stack we have scored: audio-native metrics, OpenTelemetry tracing, a hosted MCP server and an A2A connection type, all from a 10-person company that has raised $31M and grades other people's agents without publishing any validation of its own judges.

Drata

8,500 organisations and the only inline agent kill switch in GRC: Drata's MCP Proxy evaluates a tool call before it executes, where every competitor still dashboards it afterwards. Read the fine print though. It is Limited Availability, Anthropic-only, and an anonymised slice of your data trains a model shared with other customers.

Sana

The most open enterprise knowledge agent we have scored: real model choice, 100+ connectors, an MCP client and a genuinely usable free tier. It is also now a Workday subsidiary with no MCP server of its own, no eval surface, and a shared-integration permission model that can show people files they cannot open in SharePoint.