Bookmarks

AI Stack Radar — September 11, 2026

Agents & Orchestration

Hyperagent

Hyperagent, built by Airtable founder Howie Liu, hands a task brief to an agent that gets its own sandboxed browser, filesystem, and code runtime, then ships a finished site, deck, or dashboard and keeps it current as the underlying data changes. **Replaces:** a blend of research assistant, junior ops hire, and one-off Zapier-style automations for recurring deliverable work. **Pricing:** pay-as-you-go credits, no flat subscription; example completed jobs on the site ran $3.88–$24.79.

E2B

Booting a fresh Linux microVM in about 150 milliseconds is E2B's whole trick: Firecracker-based sandboxes that give a coding or data agent its own filesystem, code execution, and network access, and that can be paused, forked, or resumed from a snapshot instead of starting cold every time. **Replaces:** rolling your own Docker or Kubernetes sandbox fleet for letting agents run untrusted code safely. **Pricing:** free Hobby tier (100 sandbox-hours/month plus a $100 signup credit), Pro at $150/month, usage billed per second beyond that.

Browserbase

When an agent needs to log in, click past a CAPTCHA, or fill out a form instead of calling a clean API, Browserbase gives it an actual, remotely controlled Chrome session to do it in. It also bundles a Search API and a Fetch API that turn any URL into markdown or JSON on request. **Replaces:** self-hosted Playwright or Puppeteer clusters for any agent that needs to act on the live web rather than call APIs. **Pricing:** free tier to try it, Developer at $20/month and Startup at $99/month, custom Scale plan above that.

Trigger.dev

Crash mid-task or get redeployed, and most agent loops just lose their place. Trigger.dev is a TypeScript runtime built to survive that: tool calls and human-approval steps persist through the failure, with retries, cron scheduling, and queues built in, and the whole thing is Apache-2.0 licensed if you'd rather run it yourself. **Replaces:** hand-rolled queue-and-worker infrastructure or a Temporal-style durable execution setup for agentic TypeScript apps. **Pricing:** free tier with $5/month included usage, $10 and $50/month paid tiers, usage-based beyond that.

MCP & Tool Plumbing

Glama

Glama indexes more than 85,000 open-source MCP servers and 20,000 connectors, lets you test one in an in-browser Inspector before installing anything, and offers one-click hosting plus a managed gateway that logs every call and controls which agents can reach which tool. **Replaces:** manually cloning and running MCP servers locally, or building your own gateway to log and govern MCP traffic. **Pricing:** browsing and hosted connectors are free; self-hosting a server on Glama's own infrastructure is paid (no published rate).

Gram

Gram, Speakeasy's MCP cloud, takes an existing API (or a plain TypeScript function) and turns it into a curated MCP toolset with its own hosted URL, handling the OAuth token refresh that would otherwise be your problem. **Replaces:** writing and deploying a custom MCP wrapper around an OpenAPI spec, plumbing included. **Pricing:** usage-based, with the first 1,000 tool calls a month free.

MintMCP

Every engineer running their own unmanaged MCP config with separately stored credentials is a governance problem waiting to surface. MintMCP fronts an organization's MCP servers with a single gateway that adds role-based access control, audit trails, and SOC 2 Type II compliance, and its Agent Monitor shows exactly which tool calls Claude or Cursor made on a given day. **Replaces:** the audit gap that comes with letting every developer configure MCP access on their own. **Pricing:** free trial, custom pricing on request for teams past that.

Coding & Dev

Qodo

Instead of reviewing just the diff in front of it, Qodo checks a pull request against a graph of the whole repository, which is how it catches a breaking change three files away; a new Agentic Toolbox now exposes that same review and standards-enforcement logic to other coding agents over MCP. **Replaces:** rule-based linting tools like SonarQube for pattern-level review, sitting next to whatever agent wrote the diff. **Pricing:** 14-day free trial, Pro Team billed per credit (roughly $0.012 each), Enterprise by quote.

Muse Code

Meta's Muse Code left beta on September 1, and its standout feature is inter-session messaging: several sub-agents can now coordinate on different parts of a large refactor at once, with a session-rewind feature that can roll both the code and the conversation back to an earlier point if one of them goes sideways. **Replaces:** a slice of what Claude Code or Codex CLI do for multi-file refactors that benefit from splitting the work across agents. **Pricing:** subscription tiers at $5, $15, and $50/month, plus pay-as-you-go API pricing.

Factory.ai

Plan, write, test, review: Factory's Droid agent runs that whole loop the same way whether you invoke it from a CLI, a desktop app, or a headless `droid exec` call inside CI, though it still stops for a human to approve the merge. **Replaces:** a scripted CI bot or a junior engineer handling routine implementation and review work. **Pricing:** self-serve monthly plans from Pro at $20 up to Max at $200, plus a Teams tier at $60 base + $40/seat; Business and Enterprise are custom.

Codeflash

Codeflash hunts through a codebase for functions worth speeding up, rewrites them, checks with benchmarks that behavior didn't change, and opens a pull request with the before-and-after numbers attached; it also runs continuously against new PRs from Claude Code or Cursor to catch a performance regression before it ships. **Replaces:** the manual profiling and optimization work a senior engineer would otherwise do by hand. **Pricing:** free tier (25 optimization credits/month, public repos only), Pro at $20/user/month for private repos, Enterprise by quote.

Text & Reasoning APIs

Router

Point an existing OpenAI or Anthropic client at Router and it silently swaps in whichever model clears your quality bar for the least money, failing over automatically if a provider has an outage; it started as internal routing infrastructure at the corporate-card company Ramp before spinning out on its own. **Replaces:** OpenRouter- or Portkey-style manual multi-provider routing and hand-rolled fallback logic. **Pricing:** the routing layer is free through the end of 2026, plus a $26 signup credit; you still pay standard token prices to whichever model it picks.

Parasail

Parasail pools GPU capacity across roughly 40 data centers and puts 39-plus open and frontier models behind one OpenAI-compatible endpoint, aimed at teams that would rather commit to a spend number than rent a fixed block of GPUs from a hyperscaler and hope they sized it right. **Replaces:** renting dedicated GPU instances when a fine-tuned or open-weight model needs to go into production fast. **Pricing:** pay-per-token, no published free tier, self-serve signup with no sales call.

Requesty

More than 600 models from over 30 providers sit behind Requesty's single endpoint, with automatic failover and response caching built in, and instead of a subscription tier to pick, it charges a flat 5% markup over whatever the underlying provider bills. **Replaces:** integrating multiple model SDKs by hand just to get provider redundancy. **Pricing:** free tier (200 requests/day on free models), otherwise the flat 5% markup with all routing and observability features included.

Novita AI

A newly released open-weight model like Qwen3.8-Max tends to show up on Novita within a day or two of launch. The platform hosts more than 200 models behind OpenAI-compatible endpoints and works out as a cheaper landing spot than standing up the deployment yourself. **Replaces:** Together AI or Fireworks as the day-one serving layer for a fresh open-weight release. **Pricing:** free to start with some models free for dev and testing, paid usage billed per token from there.

Video & Motion

Sync

Sync re-times a speaker's mouth to match new or dubbed audio at a level meant for actual film and ad work, not a quick meme clip, and it's built to hold up through occluded faces, multiple speakers, and 4K ProRes footage: exactly the conditions that break most consumer lipsync tools. **Replaces:** a manual ADR or dubbing pipeline for localization-grade work, available via API, a web studio, or Adobe Premiere and ComfyUI plugins. **Pricing:** tiered plans (Lite, Assistant, Advanced) with API access; an enterprise tier for teams needing dedicated support.

Hedra

A single photo and an audio track go in; a character that talks and emotes on cue comes out of Hedra, and its newer Hedra Agent mode chains that together with separate image, video, and audio models so a whole short can come from one workflow instead of five. **Replaces:** a talking-avatar or greenscreen presenter setup when a photo-driven character is enough to deliver the script. **Pricing:** free plan available, paid credit plans starting around $15/month, separate API pricing for developers.

Higgsfield

Instead of juggling separate accounts for Seedance, Sora, and Kling, Higgsfield puts one interface and API, called PixelFlow, in front of several third-party video models and layers its own camera-motion controls and a motion-transfer tool called Genjutsu on top. **Replaces:** managing individual accounts and APIs for each underlying video model. **Pricing:** subscriptions starting at $9/month, with per-generation credit pricing (roughly $0.16–$0.70 per image-to-video output) on top.

Voice & Audio

Speko

Speko benchmarks speech-to-text, language models, and text-to-speech across providers and routes each request to whichever wins on latency or cost for that specific language, something like an OpenRouter for voice. It's built by the founder of Hupo, whose enterprise voice AI already runs at Morgan Stanley and HSBC. **Replaces:** manually wiring and maintaining separate Deepgram- or ElevenLabs-style integrations per language or use case. **Pricing:** usage-based with transparent, pass-through provider rates ($0.0010–$0.0170/minute for STT tiers); sign up for an API key directly.

Deepgram Flux TTS

Deepgram's Flux model reads the whole conversation before it speaks a line, not just the sentence it's given, so pacing and tone stay consistent across a call, and it can handle being interrupted mid-sentence without the SSML tagging most TTS engines need to fake that. **Replaces:** bolt-on TTS engines in voice-agent stacks that need manual prompt or SSML tuning to sound natural mid-conversation. **Pricing:** free through September 12, 2026 (up to 45 concurrent streams), standard metered pricing after that.

Rime AI

Rime, founded by linguists rather than career ML researchers, offers 600-plus voices across 50-plus languages tuned for the sub-200-millisecond latency a live phone call needs, and it'll run on-prem for healthcare and finance customers who can't send call audio to a third-party cloud. **Replaces:** general-purpose TTS APIs when latency or on-prem compliance is the deciding factor. **Pricing:** free trial via self-serve signup; no public price sheet, enterprise path available separately.

Inworld

Inworld's Realtime TTS-2 listens to the actual audio of earlier turns in a conversation, not just the transcript, so it can track a shifting emotional tone, and it takes plain-language voice direction like "tired but warm" instead of SSML tags, with the same voice identity holding even if the conversation switches languages mid-sentence. **Replaces:** stateless TTS turns in voice-agent pipelines that lose conversational tone between turns. **Pricing:** metered per character (roughly 1,000 characters/minute), free signup and playground to start.

Data & Training

Lightly

LightlyStudio uses embeddings and similarity search to figure out which frames in a pile of raw video or images are actually worth labeling before any of it reaches a human annotator or a vendor like Scale; a companion product, LightlyTrain, handles self-supervised pretraining on whatever gets curated. **Replaces:** manually eyeballing raw image or video dumps, or writing custom dedup scripts, before sending frames out for labeling. **Pricing:** self-serve free tier; paid plans not publicly listed, with an enterprise demo option available.

Tinker

Write an actual training loop yourself (SFT, DPO, or RL) in Python on a laptop, and Tinker, built by Thinking Machines Lab, runs the distributed execution on managed GPUs behind the scenes using LoRA, so the algorithmic decisions stay yours even though you never touch a GPU cluster directly. **Replaces:** standing up a multi-GPU cluster with Axolotl or torchtune, or reaching for a closed provider's fine-tuning API when full control over the training loop matters. **Pricing:** usage-based per million tokens plus $0.10/GB-month for checkpoint storage; no free tier, but self-serve signup with no sales call.

Unstructured

Scanned contracts, nested tables, and handwritten forms are exactly the file types that break most parsers, and exactly what Unstructured is built to handle: it turns any of 65-plus formats into clean, structured chunks a training pipeline or a RAG system can use instead of custom OCR glue code. **Replaces:** hand-rolled PDF and HTML parsing scripts teams write to turn raw documents into usable training data. **Pricing:** free tier to try it, paid self-serve tiers above that, enterprise deployment requires sales contact.

Automation & Integration

BetterClaw

BetterClaw takes a task described in plain chat and turns it into an agent that runs on its own schedule and reports back in Slack, Telegram, Discord, or Gmail. It's built for the crowd currently paying $200 to $500 a month for a Docker container and a VPS just to keep a scheduled agent alive. **Replaces:** self-hosted agent deployments that need their own infrastructure to babysit. **Pricing:** no free tier; Basic at $19/month, Pro at $49/month, Business at $149/month, Enterprise by quote.

Timbal

A workflow builder, an agent framework, a front-end layer, and an observability tool: that's four separate purchases most teams stitch together by hand, and Timbal bundles all of it into one hosted stack with more than 100 native integrations, deployable to an EU region for teams that need the data to stay there. **Replaces:** the DIY combination of a flow builder, an agent framework, a front end, and a separate observability tool. **Pricing:** paid plans start at €25/month, usage- and seat-based, with custom enterprise pricing above that.

Lutra

Tell it what you want done, like enrich this list or summarize these emails, and Lutra writes and runs actual code behind the scenes instead of chaining visual blocks on a canvas, which is the specific complaint a few of its own customer testimonials make about outgrowing n8n. **Replaces:** a visual flow-builder once a team hits its ceiling and wants inspectable, executable code instead of opaque nodes. **Pricing:** free tier with monthly refreshing credits; paid and enterprise tiers available through sales. --- *28 products this week. Found a dead link or have something to add? Reply to the email.*