AI Stack Radar — October 11, 2026

30 hosted AI products worth a look this week, sorted by the job they do. Heavy on MCP gateways, RL environments and the plumbing under agents.

🔌 MCP & Tool Plumbing

Obot

An open-source MCP gateway that puts every approved server behind one endpoint, with SSO, per-role access, audit logs and a scan for MCP use nobody signed off on. The latest release adds "virtual" servers: one URL that bundles tools from several catalog entries, pinned to a snapshot so a catalog edit can't break a running team. Replaces: hand-edited MCP configs on every laptop. Pricing: free self-hosted Community edition up to 100 users; Cloud has a free trial, Enterprise is quoted.

Manufact

The company formerly called mcp-use. It pairs its SDK with a cloud that redeploys your MCP server on every git push and hands each branch a preview URL. The Inspector lets you poke at a server without burning model tokens, and you can check the same server against ChatGPT, Claude and Gemini clients before publishing. Replaces: a Vercel or Render deploy you rigged up for an MCP server. Pricing: free with $5 of credit and 30k requests a month; Hobby $25/month.

Metorial

Gives any agent, whether that's Claude, Cursor or something you built, a single "Magic MCP" URL that fans out to 1,000+ integrations, with no per-service tokens to juggle. Each employee gets a portal showing only the tools they're cleared for, and a filter blocks risky tool calls before they run. Replaces: a pile of self-hosted MCP servers; sits in the same lane as Composio. Pricing: free Dev tier with 500K tool calls a month; Scale $250/month.

StackOne

540-odd connectors for HR, recruiting, CRM and finance systems, callable over MCP, A2A or plain REST. Older unified APIs squash every app into one shared schema; StackOne keeps each system's own data model, and it checks tool responses for prompt injection before the agent reads them. Replaces: Merge-style unified APIs for agent work. Pricing: free Starter (1,000 credits per seat a month, no card); Team $600/month.

🤖 Agents & Orchestration

Rivet

Runs each agent or session as an "Actor": a small process with its own SQLite database, rather than a container. Idle actors hibernate for free and wake in about 12 ms with their state intact, so one durable process per user is cheap enough to be the default. Replaces: Cloudflare Durable Objects or a Temporal-plus-Redis setup. Pricing: free up to 100K awake actor-hours a month; Hobby $20/month plus usage; Apache 2.0 if you'd rather host it.

Agno

A Python agent SDK plus AgentOS, a runtime that handles sessions, memory, evals, approvals and scheduling in your own cloud. The control plane talks from your browser straight to your runtime, so conversation data never passes through Agno's servers. Role-based access landed this month. Replaces: LangGraph Platform. Pricing: free for local AgentOS and the control plane; Pro $150/month.

OpenWork

A desktop agent app for Mac, Windows and Linux that works on files where they already sit, using whichever of 50+ model providers you point it at. Teams can share skills and MCP servers in one click. Automations you build on the desktop run on your own machine; ones built on the web run in the cloud. Replaces: Claude Cowork for people who want an open-source version. Pricing: desktop app free; Cloud Team free for the first 5 seats, then $10 per seat a month.

💻 Coding & Dev

cubic

Reviews pull requests against the whole codebase and gives each one a 1–5 merge-confidence score, so the low-risk ones can be auto-approved. Overnight it sweeps the full repo for bugs that already shipped, and it learns your team's rules from senior engineers' old review comments. Replaces: CodeRabbit or Copilot code review. Pricing: free for 20 PR reviews a month and for open source; Team $30 per developer a month.

Tembo

Leaves the choice of coding agent to you. You keep Claude Code, Codex or Cursor, and Tembo runs it in a full cloud VM with databases, a terminal and a browser, triggered from a Linear ticket, a Sentry alert or a Slack mention. Running services get public preview URLs, so reviewers click through a working app instead of reading a diff. Replaces: Cursor background agents or Devin. Pricing: free trial; Pro $60/month with a usage allowance, VMs billed per second from $0.11 an hour.

Momentic

End-to-end tests written as plain-English YAML that lives in your repo. When the UI changes, a test heals itself and proposes the fix as a diff for you to approve, rather than quietly rewriting itself. September brought hosted browsers and Mo, an agent that explores the app and files what it finds; coding agents can drive it over MCP. Replaces: a hand-maintained Playwright or Cypress suite. Pricing: free for 2,000 credits a month; pay-as-you-go from $125 a month, no seat fees.

🔎 Research & Retrieval

Ceramic

A search API over its own index of 40-billion-plus web pages, run by a former Google search VP. It matches on keywords only, no embeddings, which keeps answers around 50 ms and makes it obvious why a page came back. Cloudflare picked it as a launch provider for its web search beta this month. Replaces: Brave Search API or Tavily. Pricing: 1,000 free queries with no card; pay-as-you-go after that, Pro $99/month.

Datalab

Hosted document parsing from the people who wrote Marker and Surya. You get markdown or JSON with word-level bounding boxes, schema extraction, and endpoints that split merged PDFs. This month's releases cover chart reading and a public OCR benchmark; September's model reads tracked changes and redlines in scanned contracts. Replaces: LlamaParse. Pricing: free monthly credit; conversion from $4 per 1,000 pages.

Mixedbread

Managed retrieval over your own PDFs, slides, video, code and Slack, built on Mixedbread's own embedding, reranking and vector-store models, which were trained to work together. The new Toast 1 agent breaks a question into sub-queries and returns chunks cited to the page and bounding box. Replaces: a Pinecone-plus-Cohere-Rerank pipeline you glued together yourself. Pricing: free with $5 credit; Scale $20/month; search $4 per 1,000 queries.

Context.dev

Formerly Brand.dev, now in YC's S26 batch. One API key covers scraping to markdown, crawling, structured extraction, search and page-change monitoring, with JavaScript rendering and anti-bot handling switched on by default. From the Brand.dev days it can still return a company's logo, colours and fonts in the same call. Replaces: Firecrawl, plus a separate brand-enrichment API. Pricing: free 1,000 credits a month; paid $19–499/month.

🧪 Data & Training

HUD

An SDK and cloud for building reinforcement-learning environments and running agents across thousands of copies in parallel. Built-in QA agents read the traces to catch the problems that ruin a training run: graders passing wrong answers, models gaming the reward. You can also sell finished environments to labs on HUD's marketplace. Replaces: a gym you built in-house. Pricing: SDK free; cloud $0.10 per environment-hour with $10 of credit ($100 for .edu addresses).

Prime Intellect

Hosted RL post-training from one CLI. You write the task as an environment, borrow from a hub of 2,500+ community ones, and train. Runs share GPUs, so you pay per token rather than per GPU-hour, and the new Prime Sandboxes give each rollout its own microVM. Replaces: renting H100s to run your own RL loop. Pricing: per token (small Qwen models from $0.02 per million); spot GPUs from $0.47 an hour.

Distil Labs

Point one base URL at Distil and part of your production traffic flows through it. The captured calls become eval sets and synthetic training data, and a small model gets fine-tuned on them in about half an hour. You keep the weights. Replaces: paying frontier-model prices for a narrow, repetitive task. Pricing: two free training runs with downloadable weights; ten runs for $1,000.

📈 Evals, Observability & Guardrails

Laminar

Traces agent runs as readable transcripts and groups similar failures without you writing any rules. Laminar trained its own trace-reading model, flow-1, rather than paying a frontier model to read logs, and coding agents can query traces with SQL through its CLI or MCP server. Replaces: Langfuse, if you mostly debug multi-step agents. Pricing: free tier with 7-day retention; Starter $30/month.

Raindrop

Watches live agent traffic for problems a fixed eval would miss, like a tool getting misused or behaviour drifting after a model swap. A triage agent investigates and reports in Slack, and the new Simulations feature replays a change against past traces before you merge it. Replaces: sampling production logs by hand. Pricing: free for 1,000 events a month; Pro $299/month plus per-event charges.

Confident AI

The hosted side of the open-source DeepEval library: 50+ metrics, multi-turn conversation simulation, regression tests in CI and red-teaming against the OWASP LLM Top 10. It scores every production trace, not a sample, and a policy gate can block a deploy that drops below your bar. Replaces: Braintrust or Arize. Pricing: free for 2 seats and 5 test runs a week; Starter $200/month.

🎨 Image & Design

QuiverAI

Generates real, editable SVG from a prompt or a raster image, and can animate it, using its own Arrow 2 model rather than tracing pixels after the fact. Sandbox API keys return canned output and are never billed, so you can wire it into a build before spending anything. Replaces: Illustrator's Image Trace or Vectorizer.ai. Pricing: plans from $8/month with a 14-day trial; API billed from a prepaid balance.

Meshy

Turns text or a reference image into a textured, rigged 3D model, with 600+ motion presets and print-ready meshes. Version 7 is about matching the reference image more faithfully, and an MCP server means an agent can order up a 3D asset directly. Replaces: a first pass of Blender modelling. Pricing: free for 100 credits a month with no card; Pro adds API access.

🎬 Video & Motion

Decart

Lucy 2.5 edits video as it plays, at 30 fps and 720p. Name one thing, say the jacket or the background, and only that thing changes. Runway and Kling render a clip after you upload it; Decart works on a live stream with almost no lag. Replaces: offline video-to-video tools, or AR filters for live streams. Pricing: usage-based, a few cents per second of video; trial credits on signup.

Clueso

Starts from a raw screen recording and produces a polished product video with AI voiceover, captions and translation into 150+ languages, plus a matching help article. A new MCP server lets Claude or Cursor make and edit the videos, and saved styles keep every release video on brand. Replaces: Loom plus an editor, or Synthesia for training content. Pricing: free for 10 export minutes a month; Solo $40/month.

ChatCut

You tell it the edit in chat: cuts, captions, B-roll, music. It lays the result out on a normal multitrack timeline you can still drag around by hand. It works on your own footage rather than generating a fresh clip, and ChatGPT, Claude or Codex can drive it. Replaces: Descript or CapCut for routine edits. Pricing: free tier; paid $25–200/month by credits.

🎙️ Voice & Audio

Soniox

Speech-to-text, text-to-speech and live translation through one API and one multilingual model. You don't pick a language, and it copes when a speaker switches mid-sentence. TTS v2 added inline emotion tags and voice cloning from 20 seconds of audio, and processing can be pinned to the US, EU or Japan. Replaces: buying STT and TTS from two different vendors. Pricing: STT from $0.10 an hour; TTS about $0.70 per generated hour.

Gladia

A speech-to-text API whose Solaria-3 model is tuned for messy business audio, including 8 kHz phone lines and calls that switch languages. Speaker diarization and sentiment come at no extra charge, and there are native hooks for LiveKit, Pipecat and Twilio. Replaces: AssemblyAI or Speechmatics. Pricing: €50 of credit that never expires; pay-as-you-go from $0.61 an hour.

Smallest.ai

A real-time voice stack whose most interesting part is Hydra (beta), a full-duplex model that hears speech and talks back over one WebSocket, with no separate STT and TTS steps in between. It handles interruptions on the server and can call tools mid-conversation. Replaces: a Vapi-style STT-LLM-TTS chain, or OpenAI's Realtime API. Pricing: $10 free credit; voice agents from $0.05 a minute.

🖥️ Local & Self-Hosted Inference

Osaurus

A local-first desktop agent for reviewing contracts and other sensitive files, running its own small Raptor model or one you bring. It strips sensitive data before anything goes to a cloud model, asks before editing a file, and keeps a log of exactly what left your machine. Replaces: the ChatGPT desktop app for documents you'd rather not upload. Pricing: app free (MIT); team Workspaces $20/month for 5 seats.

Msty

Started as a local-plus-cloud chat app and now adds an admin layer that LM Studio and Jan lack. Nexus is a self-run gateway that controls which models, keys and policies a team can use, and Go runs agents whose every step you can review. The company says the apps send no product telemetry. Replaces: Open WebUI or LibreChat for a team rollout. Pricing: Studio free; Aurum $149/year; Go free in beta.

---

30 products this week. Found a dead link or have something to add? Reply to the email.