Start Mistral Vibe: 2 Web Sessions If Enabled, CLI in the Folder
On 19 Nov 2026 Vibe Code Web lists 2 free sessions a day where access is on, and 100 for paid users. The CLI starts with vibe in the folder.
Discover top-rated Coding & Dev applications tested for speed, accuracy, and return on investment. View side-by-side specs and verified field reviews below.
| Tool | Pricing Model | Free Tier / Entry Cost | Best For | Rating | Review |
|---|---|---|---|---|---|
| Magpie | Free | Homepage read Nov 19, 2026 says Free, open source, macOS 12 and later, Intel Mac, Windows, and Linux. GitHub lists the license as MIT. The README prints no dollar price for the app. A key you paste can belong to a paid provider plan. | People who want one local install to point Claude Code or Codex at a provider key they already hold. | ★ 4.4 | Details → |
| Cloudflare Workers AI | Freemium | Free and Paid plans each include 10,000 Neurons a day, resetting at 00:00 UTC. Free calls fail after that. Workers Paid bills $0.011 per 1,000 Neurons above it. Pricing page updated Nov 12, 2026; read Nov 16, 2026. | Developers who want a first model call on Cloudflare from a REST request. | ★ 4.5 | Details → |
| Ollama | Freemium | Local runs on your own hardware are unlimited. Cloud is billed per token. Pay as you go has no monthly fee. Pro is $20 a month and includes $60 of credits. Read Nov 16, 2026. | People who want a model on their own computer and a local API an app can call. | ★ 4.6 | Details → |
| Groq | Freemium | Free tier: gpt-oss-120b and 20b at 30 RPM, 1,000 requests/day, 8,000 TPM, 200K tokens/day. Developer: $0.15/1M input, 1,000 RPM, 250K TPM. | Developers building low-latency chat or Whisper transcription prototypes who want a free API tier before scaling to paid throughput. | ★ 4.5 | Details → |
| Gemini API | Freemium | Free tier covers Hydrogen, 3.8/3.6 Flash, and Flash-Lite; Paid Hydrogen is $1.50/1M input and $7.50/1M output. Carbon and Argon are paid-only ($2/$10, rising to $4/$20 on Dec 1). | Teams prototyping multimodal apps on Hydrogen or Flash for free, or production workloads needing data privacy and Carbon/Argon reasoning. | ★ 4.6 | Details → |
| OpenRouter | Freemium | :free model IDs allow 20 RPM and 50 req/day (raised to 1,000 req/day once 9-10+ lifetime credits are purchased); paid models bill per token with no platform cap. | Developers benchmarking multiple model vendors through one API, needing automatic provider failover, or testing :free variants. | ★ 4.4 | Details → |
| Kokonut UI | Freemium | Open-source MIT registry is free; Kokonut UI Pro is a one-time purchase of $119 (1 seat) or $374 (up to 10 seats) for 150+ components and 6 templates, no subscription. | React and Next.js developers building AI interfaces who want editable source code via shadcn CLI and optional one-time Pro templates. | ★ 4.2 | Details → |
| Antelog | Freemium | Free personal plan is $0 (was $20) with local rule blocking, 50,000 AI checks/month per machine, and 14-day history; Team plan is custom pricing. | Individual developers running Claude Code or Cursor locally who want zero-cost protection against accidental secret leaks or data loss. | ★ 4.3 | Details → |
| Burnstate | Paid | Direct download, from the FAQ on burnstate.dev verified Nov 14, 2026: a one-time license with updates, $49.99 for 1 Mac, $69.99 for 2 Macs, $89.99 for 3 Macs. The same license is also sold as a subscription from $3.99 a month or $24.99 a year. Creem emails the receipt and the key. Setapp includes Burnstate with 260+ other Mac apps and its own 7-day trial; the Setapp membership price is not printed on burnstate.dev. The Mac App Store edition is described as coming soon, a one-time purchase inside Apple's sandbox. JSON-LD on the page lists an aggregate offer from $49.99 to $89.99. | A Mac owner on macOS 15 who pays for Claude Code and Codex and wants one menu bar view of both windows, at $49.99 once for a single Mac. | ★ 4.3 | Details → |
| Plurai | Freemium | Simulation plans on plurai.ai/pricing, verified Nov 14, 2026. Free is $0 and one-time: 12,300 credits, 2 agent setups, 7 scenarios, 3 runs, 10 sandbox hours, no pay-as-you-go. Base is $60 a month: 28,710 credits every month, 3 agent setups, 14 scenarios, 20 runs a month, 20 traces analyzed a month, 30 sandbox hours a month, then $0.02 per credit. Business is $210 a month: 65,570 credits every month, 3 agent setups or 1 setup with 16 extra scenarios, 28 scenarios or 44 with a single setup, 60 runs a month, 40 traces a month, 100 sandbox hours a month, then $0.015 per credit. The page defines a Sim credit as one cent of model-provider usage at list price, so the free grant is $123 of list-price usage and the Base grant is $287.10. Overage is off until an admin sets a limit that starts at $0; past the included amount, runs stop. Free never charges. Post-training is a separate tab on the same page and is not included in these three prices. Cursor was marked Soon on the homepage install list the same day. | A developer who already runs Claude Code, Copilot CLI or Codex and wants 3 free simulation runs before paying $60 a month for 20. | ★ 4.1 | Details → |
| offstage cloud | Paid | Usage pricing only: $1.10 an hour for each sandbox or CI task, billed per second, rounded up to the whole second, with no minimum, so a 10-minute run costs about $0.18. A sandbox bills from the moment a Mac takes it until it closes and a task from start to finish; waiting in the queue is free. New personal accounts get $5 in free credit with no credit card, then add card credit in the console or turn on auto-reload. Every new team starts with a $50 monthly budget with alerts at 50%, 80% and 100%, which the owner can raise or lift. Creating a task reserves its worst case: timeoutMs converted to hours times $1.10 an hour. For comparison the site lists GitHub Actions at $3.72/hour, Namespace at $2.40/hour prepaid, Blacksmith at $4.80/hour, and Scaleway, AWS and Daytona on 24-hour minimums (offstage.cloud homepage and /docs Pricing and limits, prices checked Nov 7, 2026; verified Nov 13, 2026). | Teams whose agents or CI pipelines must build iOS apps, run Xcode or Swift tests, or drive a Mac desktop, and who would rather pay $0.18 for a 10-minute job than rent a Mac for a full day. | ★ 4.4 | Details → |
| three.dev | Freemium | Free $0 a month: room for about 50,000 requests a month at 5k tokens each, 1 GB of data a month, 5,000 requests scored a month, 30 days of history, unlimited seats. Starter $79 a month: about 250,000 requests of room, 5 GB of data a month then $3 per GB, 10,000 scored then $2.50 per 1,000, 90 days of history. Pro $249 a month: about 1,000,000 requests of room, 20 GB then $2 per GB, 25,000 scored then $1.50 per 1,000, 180 days of history. Enterprise is custom. Scoring and unlimited seats are on every plan, and the launch offer gives six months of Starter free to teams signing up by November 22, 2026; prices are in US dollars (three.dev/pricing, verified Nov 13, 2026). | Teams shipping LLM agents, support bots or copilots that need evidence of what fails in production and a way to route each failure mode straight into their coding agent. | ★ 4.5 | Details → |
| Blendra | Freemium | Free at $0: 1 project, 5 recordings, 10 scenarios, 10 checks, 5 API tests, 5 AI requests a month (each press of an AI button that explains a failure in plain English uses one; AI never writes checks or code), Playwright TypeScript export and 60 cloud runner minutes a month (up to 5 minutes per run), no credit card. Pro $79/month: 1 project, CLI and API tokens, 2,000 minutes. Pro+ $119/month: 3 projects at +$99/month per extra project, adds Playwright Python and Cypress, 5,000 minutes. Team $399/month: 5 seats at +$50/month per extra seat, roles and invites, 15,000 shared minutes. A 7-day Pro trial needs no card and nothing is deleted when it ends. Payments by Paddle as merchant of record; the pricing page publishes these detailed caps for the Free tier only; the recorder extension is on the Chrome Web Store and the npx blendra CLI is MIT on npm (blendra.app homepage and Pricing section, verified Nov 13, 2026). | QA engineers and developers who want recorded flows to land in their repo as real Playwright or Cypress code with CI wiring, instead of a hosted recorder they have to keep paying to run. | ★ 4.4 | Details → |
| Foxpit | Paid | One plan: Pro at $5 for the first month, then $12/month, billed through Polar, cancel anytime, no free trial and no free tier (foxpit.dev homepage, Pricing and FAQ sections, verified Nov 13, 2026). Model access always comes from your own accounts: the CLI agents (Claude Code, Codex, OpenCode or Antigravity) run on your own subscriptions, or the built-in Foxpit engine connects to your own provider API keys or a local server such as Ollama or LM Studio (the built-in engine is Windows-only for now, per the FAQ and Getting Started docs); Foxpit never resells model usage. | Developers who already pay for Claude Code, Codex or OpenCode and want one lead agent to split work across several of them, with shared project memory and automatic resume at usage limits. | ★ 4.3 | Details → |
| Skillbox | Freemium | Free plan at GBP 0/month, no card: one library, delivery across tools, source tracking and updates, Skillbox Trust security review, automatic maintenance, up to 3 devices and 100 followed sources. Individual at GBP 6.99/month: adds Context trim, Skillbox AI and Try in chat, with no device or source limits. Team at GBP 12.99 per member/month: everything in Individual plus shared workspaces, team standards, workspace usage stats, weekly report and more library storage. Card timing: the library itself stays free with no card; Trim, AI and Try in chat each start a 7-day Individual trial the moment you switch them on, and switching one on requires a card up front, with GBP 6.99/month billing after the trial unless cancelled (skillbox.so/pricing, verified Nov 12, 2026). | Developers running Claude Code and Cursor side by side who keep losing track of which skill lives where, and teams that want reviewed, standardized agent skills instead of ad-hoc copies in every repo. | ★ 4.3 | Details → |
| Qoder | Freemium | Free tier is $0: a 30-day daily check-in pays out 3,000 Credits, new accounts get a 2-week Pro trial with 300 Credits, and Bring Your Own API Key is supported. Lite $10/mo with 1,000 base Credits (first purchase adds 1,000 bonus Credits, renewals 500). Pro $20/mo with 2,000 base Credits plus 2,000 first-purchase bonus and a limited-time 8,000 SOTA bonus. Pro+ $60/mo with 6,000 base Credits. Ultra $200/mo with 20,000 base Credits. Credit Pack $20 per 1,500 Credits, valid 1 month from purchase. QoderWake cloud computer $10/mo per unit (list price $15). Prices exclude tax (qoder.com/pricing, verified Nov 11, 2026). | Developers and small teams who want long, multi-step engineering jobs handled in the background: free check-in Credits make the first week genuinely usable, and Quest mode fits repo-wide refactors and test sweeps. | ★ 4.5 | Details → |
| Cursor | Freemium | Free trial with 2,000 completions; Pro at $20/mo with 500 fast premium requests and unlimited slow requests. | Full-stack software engineers, indie hackers, and teams shipping complex software rapidly. | ★ 4.9 | Details → |
| GitHub Copilot | Paid | Individual plan at $10/mo ($100/yr); Copilot Business at $19/user/mo with enterprise policy control. | Enterprise engineering teams, developers working in JetBrains IDEs, and open-source contributors. | ★ 4.7 | Details → |
| v0 by Vercel | Freemium | Free credits renewed monthly; Premium at $20/mo with unlimited prompt generations and priority queues. | Frontend developers, product designers, and founders building modern web applications quickly. | ★ 4.8 | Details → |
| Replit Agent | Paid | Requires Replit Core subscription starting at $25/mo or Replit Teams with agent usage included. | Solo entrepreneurs, non-technical founders, and product managers rapid-prototyping full-stack web applications. | ★ 4.7 | Details → |
| Cohere Trial API Key | Freemium | Trial (evaluation) keys are free, rate limited, and not permitted for production or commercial use; production keys are billed pay-as-you-go (cohere.com/pricing, read 2026-11-15). Trial keys, and production keys on newer Chat model variants, are limited to 1,000 API calls a month, with the per-model and per-endpoint caps on docs.cohere.com/docs/rate-limits. | ★ 0 | Details → | |
| Chrome DevTools MCP | Freemium | Fully free: an official open-source project from the Google Chrome DevTools team (Apache-2.0). The npm package chrome-devtools-mcp is at 1.13.0 (registry record updated Nov 11, 2026). There is no paid tier, no usage quota and no account to create; the only requirements are Node.js LTS, npm and a current stable Chrome on your machine. | ★ 0 | Details → | |
| Hugging Face Inference Providers | Freemium | Free users have no included monthly credits and purchase credits for pay-as-you-go. PRO is $9/month with $2.00 of included compute credits each month, and Team or Enterprise organizations get $2.00 per seat. That $2.00 is a site-wide compute credit used for Inference Providers, Inference Endpoints, Spaces, and Jobs (huggingface.co/docs/inference-providers/en/pricing and huggingface.co/docs/hub/billing, read 2026-11-15). Hugging Face charges the provider rate with no added fee. A custom provider key is billed by that provider, and included credits do not apply to those calls. | ★ 0 | Details → | |
| Plannotator | Freemium | The local plugin is free and open source (Apache-2.0 or MIT, dual-licensed, at the user's choice): install with curl -fsSL https://plannotator.ai/install.sh | bash, no account and no quota. The announced cloud edition (shared plans, review history, integrations and verification standards) has no printed price; on Nov 15, 2026 the site only offered a waitlist. | ★ 0 | Details → | |
| Qwen Code | Freemium | The CLI is free and open source under Apache-2.0. Model usage is billed by the auth option you pick, and the official auth docs print no free quota. Five top-level choices: Qwen subscription OAuth (Plus or Pro on chat.qwen.ai, browser sign-in; OAuth cannot be configured with environment variables alone); QwenCloud (the recommended setup: a Token Plan with a monthly Credit quota, Personal or Team, or a standard pay-as-you-go API key; pick the China platform or International QwenCloud explicitly); Alibaba Cloud ModelStudio (submenu: Coding Plan, a ModelStudio Token Plan billed on actual tokens, kept apart from the QwenCloud Token Plan monthly Credit, plus a standard API key); third-party providers (DeepSeek, Grok, MiniMax, Z.AI, Kimi, Idealab, ModelScope, OpenRouter, Requesty; each key is billed by that provider); custom (a local server, a proxy, or an OpenAI, Anthropic, or Gemini compatible endpoint). You can use Qwen Code without a Qwen or Alibaba Cloud plan, and your own key is billed by that provider. ModelScope is one name on the third-party list. | ★ 0 | Details → | |
| Expo Simulators | Freemium | Expo pricing page (https://expo.dev/pricing), read 2026-11-15. Simulator minutes: iOS Simulator $0.10 per minute and Android Emulator $0.10 per minute, billed per second while a session runs. These rates apply to Starter and Production; Enterprise pricing is custom. Starter ($19/month + usage), Production ($199/month + usage), and Enterprise (custom) each include 60 minutes per month until 2027. Minutes past that allowance on Starter and Production bill at $0.10 per minute, billed per second. Free labels Simulator minutes Coming soon, so the Free plan currently includes no simulator minutes. iOS Simulator sessions and Android Emulator sessions start on Starter. Queue priority is High priority on Starter and above. Starter lists 5 concurrencies (up to 2 on macOS, $50 per extra, up to 5 extra). Agent control starts on Starter. Embed sessions start on Production. The Expo Free plan is $0/month and includes 15 Android builds, 15 iOS builds, 60 minutes of CI/CD Workflows, and updates for 1K MAUs. | ★ 0 | Details → |
Run the install script, add one provider key, and start a new Claude Code session on the model you set.
Magpie is a local gateway that points Claude Code, Codex, or Gemini CLI at a provider you add. On Nov 19, 2026 the GitHub API for github.com/yetone/magpie reported 38,764 stars, license MIT, and release v0.1.4516 published at 2026-11-19T06:15:27Z. The README install line is: curl -fsSL https://usemagpie.ai/install.sh | sh. The install.sh header, read the same day, says each download is checked against the release SHA-256. On macOS it puts magpie.app in /Applications, or in the home Applications folder when that directory is not writable, and links the magpie command into ~/.local/bin. On Linux it installs the desktop app when GTK 3 and WebKitGTK 4.1 are already present. If those libraries are missing and someone is at the terminal, the script's own line says it installs WebKitGTK 4.1 with the package manager, then the desktop app. It does not name a GTK 3 package in that step. With no terminal, no sudo, or a failed install, it installs the terminal build. The README says Mac builds are signed and notarised. It then says to open Providers, add a provider, pick a model on the Agents page, and start a new session. The terminal samples on that page are: magpie provider add deepseek followed by a key, magpie claude deepseek/deepseek-v4-pro, and magpie codex moonshot/kimi-k2.5. Those two model ids are the README samples. This entry did not call DeepSeek or Moonshot, and did not run the installer. The homepage says the app does not read API keys from environment variables. The README says masking stays off until you turn it on under Settings, then Privacy. The same page says that when masking is on and cannot run, the request is held back and the status on that row is 503. This entry did not send a masked request. Sponsor lines on the homepage name other companies. Those figures are not a Magpie price.
Create a Workers AI token and send one REST prompt to run Llama 3.1 8B on Cloudflare.
Open Workers AI, choose Use REST API, create a token, and copy the account ID. POST the prompt to the Llama 3.1 8B endpoint on api.cloudflare.com. That sample model sits outside the nine frontier models that need a paid plan or prepaid credits. A token you create yourself needs Workers AI Read and Workers AI Edit. Neuron usage shows in the Workers AI dashboard.
Install Ollama, pull gemma4:e2b, and chat with the model on your own computer.
Download Ollama, then run ollama pull gemma4:e2b and ollama run gemma4:e2b. Type /bye to leave the chat. The download is about 7.2 GB, and the quickstart recommends 8 GB of VRAM. Apps can POST to localhost port 11434 with no API key. On Linux, start the server with ollama serve if it is not already running. Cloud models use a separate API key.
High-speed LLM inference API powered by LPU hardware, offering free developer quotas for open-weight and speech models.
Groq provides low-latency inference APIs for open-weight chat models (gpt-oss-120b, Qwen 3.8), Llama prompt-guard models, Whisper v3 transcription, and Orpheus speech. Quotas are enforced at the organization level with cached tokens excluded, making it straightforward to prototype on the free tier before moving to the Developer plan.
Google's official multimodal developer API, offering free-tier prototyping on Gemini 4 Hydrogen and Flash models.
The Gemini Developer API exposes Google's full multimodal model lineup from lightweight Flash-Lite to flagship Carbon and Argon with built-in thinking tokens. Free-tier quotas reset daily per project at midnight Pacific time for prototyping, while paid tiers guarantee data privacy (no training use) and unlock Google Search grounding.
Unified AI model router providing single-API access to hundreds of commercial and open-weight models, including free tiers.
OpenRouter unifies multiple AI model providers behind one OpenAI-compatible API, letting teams benchmark and switch models without managing separate vendor keys. When an upstream provider throttles requests, OpenRouter automatically fails over across alternate providers, and exposes /api/v1/key for tracking daily free-model usage.
React, Tailwind, and Motion UI registry tailored for AI interfaces, installable as editable source code via shadcn CLI.
Kokonut UI provides specialized React, Tailwind, and Motion components for AI apps, including tool approval, model selectors, inline citations, and voice controls. Using the shadcn CLI or MCP server, components copy directly into your codebase for full ownership, while the one-time Pro registry adds 150+ components and 6 full templates.
Local pre-execution guardrail for Claude Code, Cursor, and Codex that blocks secret leaks and destructive commands.
Antelog runs locally on macOS, Linux, or Windows and hooks into Claude Code, Cursor, Codex, Gemini CLI, and Copilot with one command. Local rules intercept high-risk actions, like leaking API keys or wiping production databases, in milliseconds before execution, while optional AI checks add brief safety annotations.
A macOS 15 menu bar meter for Claude Code, Codex, Cursor, Vercel and Setapp AI limits, sold from $49.99 once or $3.99 a month.
Burnstate is a SwiftUI menu bar app for macOS 15 Sequoia or later. The FAQ says it tracks Claude Code, Codex, Cursor, Vercel and Setapp AI, each with its own sessions, weekly windows, billing cycles or credit balances. A white tick on each meter marks how much of that window has passed; the FAQ's own example of a window is five hours or a week. If usage is ahead of the tick, the app says when you would run out at the current pace. On launch it adds a hook to ~/.claude/settings.json so Claude Code permission prompts and multiple-choice questions can be answered from a card. The terminal keeps its own prompt, and whichever answer lands first wins. Cards do not take focus. There are seven card types, each with a switch. Cooldown reminds you after 90 minutes at the keyboard, around lunch and dinner, and at the times you set to wind down and stop. A break counts only time away, and it can lock the Mac until the break ends. Cost figures apply published per-token API prices, including cache reads and writes, to local logs. The FAQ says subscriptions are not billed per token, and Cursor's numbers come from Cursor. Limit numbers are requested from each provider's own API. The FAQ says the direct edition's call to burnstate.dev is the license check and that this call sends nothing else. Setapp AI insights leave the Mac only when you ask for one. Apple Intelligence insights need macOS 26 Tahoe on Apple silicon. A burn report packs up to a year of usage into the part of a burnstate.dev link after the #, which browsers do not send to the server.
A sandbox that runs your coding agent against mocked tools and scores the failures, with a one-time free grant of 3 runs and 12,300 credits.
Plurai Simulation reads an agent as it is built and wires a digital twin into a mocked world of services, documents and world state, then scores each run. The homepage install path is a plugin for Claude Code, GitHub Copilot CLI and OpenAI Codex; Cursor was labeled Soon on Nov 14, 2026. One billed run is a single full pass of your scenarios. Inside that pass, each scenario is played more than once to see if behavior holds. The free cap of 3 runs counts these billed runs, not each inner replay. An agent setup is that wiring step, and later runs reuse the twin until the agent changes in a large way. Nothing in the free tier bills you: when the 3 runs, 10 sandbox hours or 12,300 credits are gone, simulation stops. Base at $60 a month raises the caps to 20 runs and 30 sandbox hours, with pay-as-you-go at $0.02 a credit if an admin turns overage on. The company also sells post-training on a second tab. A cost calculator on the pricing page says its comparison depends on your inputs, so this entry does not repeat those sample savings.
On-demand macOS 27 sandboxes on Mac Studio M5 Max for coding agents and CI, at $1.10 an hour billed per second.
offstage cloud hands your code or your agent a real Mac through an API. A sandbox is a macOS 27 user account on Apple silicon that you drive one call at a time: create it, run commands in zsh -lc and get stdout, stderr, exitCode and durationMs, write and read files up to 10 MiB, then kill it. Sessions are headless for command-line work or gui for apps, screenshots and desktop goals, and gui sandboxes are recorded with a replay URL that lives 7 days. runAgent gives a gui sandbox a goal that Claude carries out on the Mac using your own Anthropic API key, and there is no click or typing API, so no screen input crosses the network. Tasks are fire-and-forget CI: clone a public repo, run one command, collect named artifacts, with failureReason telling you whether the build broke, the clone failed, it timed out or no Mac took it. Machines are Mac Studio M5 Max with 18 CPU cores and 36 GB of memory shared by up to 8 sandboxes, with no fixed allocation per sandbox, running Xcode 27.2 and the iOS 27.2 simulator, Swift and SwiftPM, Apple clang, git and make, with network access but no sudo and no Homebrew. Limits are 4 open sandboxes and 4 running CI jobs per team with 16 more queued, a 1-hour default and 24-hour maximum sandbox life, 10 idle minutes before auto-close and 60 minutes per job. The SDK is TypeScript ESM with no runtime dependencies on Node 18+, Bun and Deno, a SKILL.md lets a coding agent set itself up, and a GitHub Actions runner needs one workflow line changed.
Scores your AI agent's live traffic, groups the failures by root cause, and hands each one to your coding agent over MCP to fix.
three.dev is an evaluation product for AI agents running in production, from round3, whose founders came from Microsoft, Telefónica, Docker and Datadog. You send it the conversations and LLM calls your agent already makes, sampled at a rate you set from 1% to 100%; those requests are scored as they happen, and issues are detected automatically with the evidence attached. Issues are then grouped into failure modes by root cause, each with how often it occurs and how much it matters. The homepage example: a support assistant answered a refund question with 14 days while the policy said 30, and Wrong refund window appeared in 386 of 8,124 scored requests. Fix with your agent gives the failure mode to your coding agent over MCP, which reads everything three.dev recorded and finds the fix, in that case a refund window hardcoded in support/assistant.ts. After the fix ships you compare any two slices of traffic, week before versus week after, and see each failure mode side by side, including new ones; in the example Wrong refund window fell from 386 to 16 while a new mode, Refund policy pasted in full, appeared. You choose a sampling rate from 1% to 100%, so cost follows how much of your traffic you want scored, and a bill estimator on the pricing page computes it from requests a month, average tokens and sampling rate.
Record a flow once in desktop Chrome and export a Playwright project you own: specs, page objects, locators, test data and a GitHub Actions workflow.
Blendra turns one recording into an end-to-end test project. You install the Blendra Recorder extension in desktop Chrome and use your app normally; every click, entry and navigation is saved with a locator picked in a fixed order, test ids first, then role and name, then text, so nothing is invented. No LLM writes the code: the same recording gives the same output every time. After each step Blendra lists what your app actually showed, a message, a new URL, a heading, and the outcomes you tick become checks holding those real values; an export where no test checks an outcome is blocked. Export produces a standalone project you can commit and run without Blendra, in Playwright TypeScript on every plan, or Playwright Python and Cypress on Pro+ and Team, with specs, page objects, locator files, scenario data and .github/workflows/e2e.yml. Cloud runners execute scheduled runs (60 minutes a month free, up to 15,000 shared on Team), results come back through the bundled reporter unless you set BLENDRA_REPORTER=off, and npx blendra pull updates generated code without overwriting local edits. Blendra says it was backed by startup programs from Anthropic and AWS.
Runs Claude Code, Codex, OpenCode and Antigravity as one team, with a lead agent that splits the work and shared memory across all of them.
Foxpit is a desktop app, in public beta at v0.1.27, that orchestrates the AI coding CLIs you already subscribe to. A lead agent called Fox takes one task, splits it, and hands the pieces to teammates that can each run a different agent or model, including cheaper or free ones for research and tests. Every project gets its own room in a 3D office view (which can be turned off for a plain terminal cockpit), and a terminal grid shows up to 4 real agent shells side by side. Shared memory lives in .foxpit/memory as plain markdown with [[links]] that any agent picks up in a new session. When Claude Code hits its 5-hour usage limit, Fox parks the turn, tells you when it resets, and sends continue on its own. Each agent can work in its own git worktree, finished work arrives as delivery cards with files changed, tests passed and cost, and approvals surface as desktop notifications. Data stays local: agent events travel over 127.0.0.1 only, with no telemetry. Runs on Windows 10/11, macOS on Apple Silicon and Linux as an AppImage; Node.js is not required.
One library for your AI coding agent skills, delivered to each tool's skill folder, with every version scanned for prompt injection before it lands.
Skillbox solves the sprawl of agent skills: one copy lives on GitHub, another in a project folder, three more are installed for Cursor and missing from Claude Code. You import skills from your computer, a GitHub or GitLab URL, or the Explore catalog; the library records where each skill came from, shows when the upstream changes, compares versions and lets you decide what updates. Activating a skill delivers it to connected tools on your machines: the docs wire it to ~/.claude/skills for Claude Code and ~/.cursor/skills for Cursor (Codex is named in the Product Hunt launch). Skillbox Trust reviews instructions and bundled files for prompt injection, data exfiltration, secrets and obfuscated code, labeled in STRIDE and OWASP terms, with Cisco Skill Scanner findings covering public and shared skills. Automatic maintenance here means finding overlapping skills and keeping required ones in place; upstream updates themselves stay opt-in. Collections group skills by workflow, and project scoping sends each repo only the subset that belongs to it, so a frontend project never loads a release checklist. A desktop app ships for macOS. Device caps, source caps, prices and card timing were all checked on the official site on Nov 12, 2026.
Alibaba's agentic coding platform: Plan breaks down the approach, Goal keeps long tasks on track, and Quest mode runs multi-hour jobs in the background while you keep working.
Qoder is Alibaba's agentic coding platform, shipping as a desktop app, IDE, CLI and JetBrains plugin for Windows, macOS and Linux (vendor claims 6 million users and 100,000 enterprise customers). You describe a goal in natural language; the platform drafts a Plan, executes with model routing that picks the model per task, and validates output against a Goal until delivery. Quest mode parks long jobs such as dependency upgrades, batch refactors or test runs in the background and pushes results when done. Context is built from the whole repo, docs and external API references, and a Memory layer carries personal style and team conventions across projects. The ecosystem counts 40+ connectors and 20,000+ skills over MCP, Qoder Sites publishes dynamic web apps with a backend in one sentence, and the built-in Sonus model adds desktop and computer operation for tasks beyond code. Enterprise controls cover member management, knowledge bases and audit logs.
The next-generation AI code editor built on VS Code with repo-wide context and Composer multi-file edits.
A full VS Code fork engineered from scratch for AI pairing. Features whole-repository indexing, terminal error debugging, and the breakthrough 'Composer' mode for multi-file architectural changes.
The battle-tested enterprise coding assistant supporting every major IDE and code platform.
Powered by OpenAI and GitHub. Deeply integrated into VS Code, Visual Studio, JetBrains, and Neovim, offering reliable autocomplete, PR review automation, and chat assistants.
Generative frontend engineering tool turning natural language prompts into production React & Tailwind UI.
Vercel's generative UI system. You describe a web component or full page, and v0 outputs clean, copy-pasteable React code styling with Tailwind CSS and shadcn/ui, ready to deploy.
Autonomous full-stack development agent that builds, tests, and deploys entire cloud apps from an idea.
An autonomous AI software engineer in the cloud. Give it an idea in plain text, and Replit Agent sets up dependencies, designs databases, writes backend/frontend code, and deploys it live.
Cohere Trial API Key calls Chat, Embed, Rerank, and Parse: create an account, copy the key on the API Keys page, and paste it into a quickstart.
A Cohere trial key is how you call Chat, Embed, Rerank, and Parse while you evaluate a prototype. Three copyable paths are text generation, semantic search over your own documents, and reranking of search results. Create an account at https://dashboard.cohere.com/welcome/register. The privacy policy describes a trial user as someone on a free trial account with no payment method on file. After signup, open API Keys at https://dashboard.cohere.com/api-keys and copy the trial key created with the account. Install an SDK from the installation page, then paste the key into the text-generation, semantic-search, or rerank quickstart.
Google officially hands Chrome DevTools to coding agents: an MCP server and a CLI, fully free and open source, with 56,435 GitHub stars.
chrome-devtools-mcp is the official project from the Google Chrome DevTools team that hands a live Chrome browser to coding agents (the README names Antigravity, Claude, Cursor and Copilot) for control and inspection. As of Nov 15, 2026 the repo shows 56,435 GitHub stars, was created Sep 11, 2025 and had commits that day. It ships as an MCP server and as a CLI: add an npx -y chrome-devtools-mcp@latest entry to your MCP client config (the @latest tag always runs the newest version), or install it with npm i chrome-devtools-mcp@latest -g and drive the browser from the terminal with chrome-devtools navigate_page 1 --url ..., chrome-devtools take_screenshot 1 --filePath screenshot.png and chrome-devtools stop. Three capability blocks: performance (record traces and extract actionable insights), debugging (network request analysis, screenshots and browser console messages with source-mapped stack traces), and automation (puppeteer-driven with automatic waiting for action results). Four things to know before you enable it: usage statistics are on by default and report tool invocation success rates, latency and environment information to Google; turn them off with the --no-usage-statistics flag or the CHROME_DEVTOOLS_MCP_NO_USAGE_STATISTICS or CI environment variables. The performance tools may send trace URLs to the Google CrUX API for real-user experience data; disable that with --no-performance-crux. Only Google Chrome and Chrome for Testing are officially supported, so other Chromium-based browsers are not guaranteed. And the browser does not start with the server; it launches when an agent first calls a tool that needs it. The server also checks the npm registry for updates by default, which CHROME_DEVTOOLS_MCP_NO_UPDATE_CHECKS turns off. The README's first test prompt is Check the performance of https://developers.chrome.com, and a recorded trace means your setup works.
Hugging Face Inference Providers calls chat, image, and embedding models through one token: create an account, enable the Inference Providers permission on Access Tokens, then copy the Getting Started quickstart.
Inference Providers is the Hugging Face router that calls models from partners such as Groq, Cerebras, Fireworks, Together, and Fal with one token. The overview lists three jobs: text generation for chat, content, or code help; image and video generation; and embeddings for search or RAG. Create a free account at huggingface.co. Then open Access Tokens at https://hf.co/settings/tokens and create a fine-grained token with the Make calls to Inference Providers permission. Routed calls on a free account need purchased credits. The billing doc says these compute services are paid by credit card, added at https://huggingface.co/settings/billing/payment. Copy the Quick Start on the Getting Started page (InferenceClient, or an OpenAI client with base_url https://router.huggingface.co/v1). You can try prompts first at https://huggingface.co/playground, or save your own provider key in account settings so that provider bills the call.
A local review layer between you and your AI coding agents: one Inbox collects every agent's questions, you annotate and hit Send, and the reply returns to the original session. Free and open source.
Plannotator is a local, browser-based review surface for AI coding agents, dual-licensed under Apache-2.0 or MIT, at the user's choice. On Nov 15, 2026 the repo showed 11,307 GitHub stars. The repository was created on Dec 28, 2025. The GitHub Trending daily listing is dated Nov 15, 2026. Commits were present on Nov 15, 2026. The README lists supported agents: Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode, Kiro, Droid, Amp, DeepSeek Harness and Pi, and says the Inbox also works with personal assistants (Grok Bot, Muse, Hermes, OpenClaw). Three pieces hold it together. The Plannotator Inbox is one local window where agents leave messages, questions and guided reviews while they keep working; you answer a question with a click, annotate a file and press Send, and the reply goes back to the session that asked. Claude Code, Pi and OpenCode connect through official plugins, other agents reach the Inbox over MCP with plannotator inbox mcp, and plannotator inbox --help is the guide an agent reads. The annotate layer covers plans, specs, messages, HTML artifacts, diffs and PRs, and your feedback is sent to the agent as its next message. Herdr Annotate brings the same review flow to the terminal inside herdr.dev for Claude Code, Codex, Pi, Copilot CLI and Droid replies. The README prints the lifecycle commands: install with curl -fsSL https://plannotator.ai/install.sh | bash (flags include --minimal, --version vX.Y.Z and --verify-attestation), uninstall with plannotator uninstall or plannotator uninstall --purge, and the data directory defaults to ~/.plannotator. When you set PLANNOTATOR_DATA_DIR yourself, ~/.local/share/plannotator is an example. XDG is only the fallback. Development mode runs claude --plugin-dir ./apps/hook. The site describes itself as runs locally, natively integrates with agents, free and open source; the cloud edition is waitlist-only as of Nov 15, 2026.
A terminal coding agent, one install command, and five entry points (terminal, desktop, qwen serve, editor plugins, and chat channels).
Qwen Code is the Alibaba Qwen team's open-source terminal coding agent, Apache-2.0 licensed, with 30,408 GitHub stars and npm release 0.27.3 as of Nov 15, 2026; the repo was created Jun 26, 2025 and had commits on Nov 15. One install command (npm install -g @qwen-code/qwen-code@latest, or brew install qwen-code on macOS and Linux) opens five interfaces: the terminal (run qwen inside your project), desktop apps for macOS, Windows and Linux, a browser UI through qwen serve --open (experimental), editor integrations for VS Code, Zed and JetBrains, and chat channels through Telegram, DingTalk, WeChat, WeCom, Feishu, QQ Bot, Email or GitHub bots (started with qwen channel start). Automation takes three routes: headless mode (qwen -p) for scripts and CI, SDKs in TypeScript, Python and Java, and a daemon (qwen serve) that shares sessions across browser, editor and SDK clients over HTTP and SSE with reconnect support. The skill system ships /review, /batch and /bugfix, /learn distills docs or code into new skills, /curator maintains the library, and opt-in Dynamic Workflows let the model write JavaScript that orchestrates parallel fan-out, verification passes and resumable runs. The first-run auth menu has five top-level choices. Qwen subscription OAuth uses a Plus or Pro plan on chat.qwen.ai through the browser, and OAuth cannot be configured with environment variables alone. QwenCloud is the recommended setup: a Token Plan with a monthly Credit quota (Personal and Team editions) or a standard pay-as-you-go API key, and you pick the China platform or International QwenCloud explicitly. The UI language leaves the endpoint unchanged. Alibaba Cloud ModelStudio opens a submenu with Coding Plan, a ModelStudio Token Plan billed on actual tokens (kept apart from the QwenCloud Token Plan monthly Credit), and a standard API key. Third-party providers are DeepSeek, Grok, MiniMax, Z.AI, Kimi, Idealab, ModelScope, OpenRouter, and Requesty, and each key is billed by that provider. Custom covers a local server, a proxy, or a compatible endpoint. You can use the CLI without a Qwen or Alibaba Cloud plan, with your own key billed by that provider. /doctor checks the setup. The project started as a fork of Google Gemini CLI v0.8.2 and has been developed independently since Qwen Code v0.1.
Run your app in a cloud iOS Simulator or Android Emulator, stream it in the browser, and drive the session from an agent.
Expo Simulators runs your app in a cloud iOS Simulator or Android Emulator, streams it in the browser, and lets an agent or a test job drive the session. You, an agent, or a CI job starts one with npx eas-cli@latest simulator --platform ios. The --type flag defaults to agent-device. The other controller session types are argent and appium, and every type includes a browser preview. The pricing page lists iOS Simulator and Android Emulator at $0.10 per minute, billed per second while a session runs. Starter ($19/month + usage), Production ($199/month + usage), and Enterprise (custom) include 60 minutes per month until 2027. Minutes past that allowance on Starter and Production use the same $0.10 per minute rate, billed per second. The Free row marks Simulator minutes as Coming soon, and iOS Simulator sessions and Android Emulator sessions begin on Starter.
On 19 Nov 2026 Vibe Code Web lists 2 free sessions a day where access is on, and 100 for paid users. The CLI starts with vibe in the folder.
Checked 19 Nov 2026: a new Cohere account already has a trial key. Chat is listed at 20 requests a minute, and the key stops at 1,000 calls a month.
On 19 Nov 2026, gemini-2.5-flash is off the Studio picker. From 1 Dec calls route to gemini-3.5-flash-lite. The quickstart sample is gemini-4-hydrogen.
On 18 Nov 2026, GET /v1/models returned four ids. The quickstart still names Meta-Llama-3.3-70B-Instruct, which was not among them.
The rate-limit FAQ says $5 follows a verified payment method and expires in 30 days. Skip adding one, and the playground and API stay inactive.
Pull gemma4:e2b in Ollama with no key, about 7.2 GB, or send openai/gpt-oss-20b to Groq. The quickstart sample names another id. Checked 17 Nov 2026.
Fetched 17 Nov 2026: up to 40 requests a minute and 10,000 a day. The llms.txt sample id was missing from the live list of 32.
Header: text in, text out, 200K context, 128K max output. A later note mentions cameras; the cells stay text. Input, output, and cache hits are free.
Meituan opened LongCat-2.5 weights under MIT on 10 Nov 2026. The API window is 1,048,576 tokens. Chat, one POST, or Claude Code.
Limits page, 15 Nov 2026: :free IDs allow 20 a minute. Day cap is 50 below 9 credits and 1,000 from 9. Listed token price is $0.
Free and Paid Workers plans both include 10,000 neurons a day. Free stops there. Paid bills $0.011 per extra 1,000. Reset is 00:00 UTC, as of 15 Nov 2026.
Free gpt-oss-120b on Groq: 30 RPM, 1,000 requests/day, 200,000 tokens/day. Developer price is $0.15 in and $0.60 out per million tokens, as of 15 Nov 2026.
NeptoLab's Cloady bills declared CPU, RAM and SSD. A free workspace covers 0.5 vCPU, 1 GB, 5 GB and 10 GB of monthly traffic. Paid presets run $3 to $120.
On November 14, 2026, Haiku 5.5 is $0.10 per million input tokens up to 100,000 prompt tokens, then $0.50 per million. Output is $0.50 or $2.50 per million.
Gemini 4 Hydrogen prices input at $1.50 per million tokens with a free API tier, adds an xhigh gear, and powers Orchestrator routing in Antigravity.
Windsurf raised Pro from $15 to $20 on March 19, 2026 and replaced credits with quotas. We compare message limits, team pricing and verdicts against Cursor Pro.
A practical comparison between Claude Fable 5.2 and Opus 5.5 across API pricing, subscription allowances, vendor benchmarks, and independent test results from Artificial Analysis.
A complete breakdown of Meta Muse Ember (muse-ember-1.0): token pricing, prompt caching, long-context rate jumps, and Artificial Analysis benchmarks for agentic execution.
Replit Agent builds, tests and deploys full-stack apps from a single prompt, while v0 turns prompts into production React UI. We compare plans, metering and capability limits using both vendors' official pricing pages, checked Nov 10, 2026.
Anthropic cut Fable 5.2 to $8 per million input tokens and made its outputs up to 30% faster. We break down the new pricing, the Mythos 5.2 restrictions, and when switching saves money.
LG AI Research opened the API of K-EXAONE 2.5, a 750B-parameter (31B active) MoE model with 512K context, at no charge. OpenAI-compatible, tool calling and structured outputs included. Here is what it is, how to plug it into your existing stack, and where the catch is.
TermiX is an agent-to-agent job marketplace: agents post funded tasks, bid on work, and settle through on-chain escrow with zkVM arbitration. Here's how it works and how to plug your own agent in.
Moonshot published the full 2.8T-parameter Kimi K3.1 Pro weights on Hugging Face on Nov 2. We break down the $2/$10 API pricing, benchmark claims and whether self-hosting makes sense.
We benchmarked the three leading AI code editors across multi-file refactoring, terminal self-healing, and price-to-performance to identify the ultimate engineering workflow.