Tooling Catalog
Browser, Sandbox, and Integration
Browser agents, web extraction, sandboxed execution, and external app tool hubs
Last verified 2026-09-25
Use these when agents need live web context, GUI automation, isolated code execution, or scoped actions across SaaS systems.
Agent Sandboxes and Compute
Use these when agents need to run code, execute untrusted workloads, or access GPUs without turning the main app server into an execution environment. For the decision matrix on choosing among them, see sandboxes and workspaces.
| Tool | Website | What it is for | Tier | Notes |
|---|---|---|---|---|
| Daytona | daytona.io | Secure, stateful for AI-generated code and agent workflows | Conditional | Hosted product active. Went closed-source June 2026, citing AI-assisted vulnerability discovery against its isolation layer; the old daytonaio/daytona repo stays public but unmaintained, with and docs moved to github.com/daytona. Nightona continues an AGPL-3.0 community fork from the last open release. |
| E2B Desktop | docs.e2b.dev | GUI desktop sandboxes for computer-use agents, built on E2B's sandbox infrastructure | Conditional | Use alongside E2B's code sandboxes when the agent needs a full desktop rather than a code interpreter, such as testing computer-use models. |
| Modal | modal.com | Serverless compute, GPUs, sandboxes, and batch jobs for AI workloads | Conditional | VM Sandboxes shipped June 2026 and integrate with Claude Managed Agents. Useful for agent tasks that need elastic Python/GPU execution. |
| Cloudflare Sandbox SDK | developers.cloudflare.com | Isolated container sandboxes for command execution, files, terminals, and code interpreter workflows from Workers | Conditional | Sandbox SDK and Containers reached GA 2026-04-13. Strong fit when code execution should be controlled by a Worker and stay close to edge-hosted agents. |
| Agentuity Sandboxes | agentuity.com | Isolated containers inside an agent deployment platform | Watch | Track as part of agent-native infrastructure rather than generic cloud hosting. |
| Vercel Sandbox | vercel.com | Ephemeral microVMs for running generated or untrusted code | Conditional | GA 2026-01-30; workspace persistence GA 2026-05-26. Managed Images (versioned, extensible base images) replaced the older runtimes 2026-08-10, and sandboxes now run in all 20 Vercel compute regions as of 2026-09-10, up from 4. Relevant when AI code execution should stay isolated from the app runtime. |
| E2B | e2b.dev | Secure sandboxes and code interpreter infrastructure for | Conditional | Positioned as an enterprise agent cloud with Secure MCPs. Strong fit for data analysis, generated-code execution, and code-interpreter features. |
| Fly Machines / Sprites | fly.io | Fast-booting microVMs (Machines) and agent-oriented Sprites for isolated code execution | Conditional | Good fit when agents need fast per-session microVMs close to a Fly-hosted app. |
| Blaxel | blaxel.ai | Agent infrastructure platform with fast sandboxes for isolated code execution | Watch | Acquired by platform Baseten, announced 2026-09-10; the two are integrating agent execution into Baseten's model-serving stack. Track how the product changes under new ownership before treating it as foundational. |
| OpenAI hosted Shell tool | platform.openai.com | Platform-attached terminal containers invoked directly from the Responses | Conditional | Fresh-per-request or reused-across-turns containers with Agent Skills loaded in; the zero-infrastructure default for OpenAI-native agents. |
| Amazon Bedrock AgentCore Code Interpreter | docs.aws.amazon.com | Managed code interpreter for Bedrock AgentCore agents | Conditional | Good fit when agent execution, browser, and code interpretation are intentionally AWS-native. |
| Azure Container Apps dynamic sessions | learn.microsoft.com | Session-based sandboxed code interpreter backed by Azure Container Apps | Conditional | Good Azure-native route for Python execution and uploaded-file analysis. |
| Riza Code Interpreter | docs.riza.io | Hosted and self-hostable code interpreter for model-generated code | Watch | Surfaced in LangChain integrations; evaluate if language coverage or self-hosting is important. |
| bash-tool | github.com/vercel-labs | AI SDK-compatible bash tool intended for sandboxed execution | Watch | Useful pattern for controlled shell execution; never expose it without a sandbox and policy layer. |
| ctx-zip | github.com/karthikscale3 | Converts and AI SDK tools into code-mode tools that run through Vercel Sandbox | Watch | Interesting for large tool sets and code-mode execution, but keep it behind strict permissions. |
| Morph | docs.morphllm.com | Fast code editing and apply-patch infrastructure for AI coding workflows | Watch | Useful for coding agents that need reliable patch application rather than raw text replacement. |
Agent Web Access and Automation
These are included because their primary use is giving agents live web context, browser control, or automation. Generic automation platforms only stay if they expose AI-agent-specific nodes or tool surfaces. For the decision matrix on when an agent needs a browser, see browser access.
Two capabilities are worth checking for beyond the feature lists below, because a browser agent is otherwise the least observable thing in a system. The first is a live view of the running session - most hosted providers expose one, and Firecrawl returns a liveViewUrl alongside the CDP endpoint for an interactive session. The second is session recording: an artifact you can replay after the run, share, or attach to a bug report. Framework-side support for this is early - Mastra's browser recording is beta, writes Motion-JPEG AVI files with optional per-step captions, and caps a recording at 120 seconds.
| Tool | Website | What it is for | Tier | Notes |
|---|---|---|---|---|
| Browserbase | docs.browserbase.com | Cloud browsers, sessions, identity, observability, and infrastructure for web agents | Core | Strong choice when agents must browse authenticated, JavaScript-heavy, or interactive sites. |
| Browser Use | docs.browser-use.com | Open-source and managed AI browser automation for task-level web agents | Conditional | Useful when a Python or hosted browser agent should plan, browse, extract, and recover across multi-step web tasks. |
| Kernel | kernel.sh | Fast browser infrastructure for web agents and browser automation | Conditional | Good signal for production browser-agent infrastructure. |
| Stagehand | docs.browserbase.com | AI-native browser automation built around Playwright plus act, extract, and observe | Core | Good bridge between deterministic browser automation and LLM-driven page interaction. Documented and maintained under Browserbase, and configured with env: "BROWSERBASE" by default - treat the two as one adoption decision unless you have verified a local-browser path. |
| Browser Run | developers.cloudflare.com | Headless browser automation with screenshots, PDFs, extraction, crawling, Playwright, CDP, Stagehand, and MCP options | Conditional | Strong fit when browser execution should be available as managed infrastructure near edge agents. |
| Amazon Bedrock AgentCore Browser | docs.aws.amazon.com | Managed browser automation capability for Bedrock AgentCore | Conditional | Use when the agent stack is already AWS/Bedrock-native and browser control needs to stay inside AWS boundaries. |
| Perplexity Personal Computer | perplexity.ai | Consumer desktop agent that acts across a user's local files, native apps, and the browser, not a developer API | Watch | Windows launch 2026-07-28, after Mac. Not an integration surface - worth tracking as a reference point for how far end-user computer-use agents have reached: approval prompts before sensitive or hard-to-reverse actions, scoped folder access. |
| Firecrawl | docs.firecrawl.dev | Search, scrape, crawl, map, interact, and hosted browser sandbox APIs with LLM-ready output and MCP support | Core | Good for docs ingestion, web research, and pipelines. Now also ships Browser Sandbox - managed browser sessions for interactive workflows, returning a CDP URL and a live view URL - so it spans extraction and session-based browser control rather than scraping alone. |
| Tavily | docs.tavily.com | Search, extract, crawl, map, and research API for AI applications | Conditional | Good for web-aware agents that need current search results. |
| Exa | exa.ai | AI-oriented web search, contents extraction, and structured web research | Conditional | Good for semantic web retrieval and research agents. |
| Parallel | parallel.ai | Web search and extraction tools for agents with token-efficient context | Watch | Surfaced by the AI SDK tool registry; evaluate where compressed web results matter. |
| Perplexity Search API | docs.perplexity.ai | Search API for current web answers and source-backed search results | Conditional | Useful when a workflow wants search-backed synthesis rather than raw crawler output. |
| Valyu | docs.valyu.ai | Domain-focused search tooling for AI agents | Watch | Surfaced by the AI SDK tool registry; evaluate for specialized research domains. |
| Airweave | docs.airweave.ai | Semantic search and RAG data-source tooling for AI apps | Watch | Useful signal for connector-backed semantic search, especially when search spans app data sources. |
| Tako Search | github.com/TakoData | Search/data/visualization tool package for AI SDK workflows | Watch | Track for data-rich research and visualization workflows. |
| Apify | docs.apify.com | Web scraping, crawling, browser automation, and actor marketplace | Conditional | Useful when agents need durable crawlers or a marketplace of maintained extractors. |
| browserless | docs.browserless.io | Hosted Chrome automation over CDP, Playwright, Puppeteer, and REST APIs | Conditional | Practical browser infrastructure when the app already owns deterministic automation code. |
| AgentQL | docs.agentql.com | Query-like web interaction and extraction for agent/browser workflows | Watch | Surfaced in LangChain browser-tool integrations; useful when selectors are brittle. |
| Hyperbrowser | docs.hyperbrowser.ai | Browser automation and web scraping infrastructure for AI agents | Watch | Appears in LangChain browser integrations; evaluate against Browserbase/Stagehand. |
| MultiOn | docs.multion.ai | Browser-agent automation API | Watch | Interesting for delegated web tasks, but validate reliability and control boundaries carefully. |
| Bright Data | docs.brightdata.com | Web data, scraping browsers, datasets, proxies, and unlocker infrastructure | Conditional | Consider for difficult extraction at scale; keep compliance and site terms explicit. |
| Oxylabs Web Scraper API | developers.oxylabs.io | Web scraping and SERP extraction infrastructure | Conditional | Consider for scale and difficult extraction where proxy/compliance operations are already understood. |
| SerpAPI | serpapi.com | Search engine results API | Conditional | Useful for search-result pages when you need structured SERP data rather than semantic search. |
| Serper | serper.dev | Google Search API for AI apps and automation | Conditional | Lightweight option for web-search tools where raw search results are enough. |
| Brave Search API | api.search.brave.com | Independent web search API | Conditional | Good alternative search source when avoiding Google-dependent SERP APIs matters. |
| Linkup | docs.linkup.so | Search and retrieval API for AI agents | Watch | Track as part of the agent-native web search category. |
| Jina Reader | jina.ai | URL-to-Markdown and web content extraction service | Conditional | Useful as a simple reader tool before adding a full crawler or browser stack. |
| Jina Search | search.jina.ai | Search API returning web results and page content for AI workflows | Conditional | Surfaced in LangChain search-tool integrations and pairs naturally with Jina Reader. |
| SearchApi | searchapi.net | Search engine results API | Conditional | Another SERP-style option; use only when its engine coverage and pricing fit. |
| SearxNG | docs.searxng.org | Self-hostable metasearch engine | Conditional | Good when search should be self-hosted or privacy-sensitive. |
| DuckDuckGo Search | duckduckgo.com | General web search source used by several toolkits | Conditional | Useful as a low-friction tool in prototypes; production use needs reliability checks. |
| You.com Search API | docs.you.com | Search API for AI applications | Watch | Appears in LangChain search integrations; evaluate against Exa/Tavily/Firecrawl/Brave. |
External App Integration and Tool Hubs
Use these when the value of the agent is acting across SaaS systems. Prefer scoped, auditable tool access over dumping every possible integration into the model context.
| Tool | Website | What it is for | Tier | Notes |
|---|---|---|---|---|
| Composio | docs.composio.dev | Managed integrations and tools for AI agents across SaaS apps | Conditional | Good for breadth and fast prototypes; keep permissions, user connection state, and explicit. |
| Arcade | docs.arcade.dev | Authenticated tools and integrations for AI agents | Watch | Useful direction for delegated app actions where auth and user consent matter. |
| Pipedream Connect | pipedream.com | User-authenticated integrations and workflows that apps can expose to AI features | Conditional | Good when agent tools need many SaaS connectors plus managed . |
| Zapier AI Actions | actions.zapier.com | Zapier actions exposed to AI assistants and agent workflows | Conditional | Good for long-tail SaaS actions, especially non-engineering workflows. |
| n8n AI Agent | docs.n8n.io | Workflow automation with AI nodes, tools, agents, and steps | Conditional | Useful for operations teams and internal workflows; use code-first runtimes for core product logic. |
| Paragon | docs.useparagon.com | Embedded integrations platform with managed auth and sync | Conditional | Useful when integrations are a product feature rather than only an agent convenience. |
| Merge | docs.merge.dev | Unified APIs for categories such as HRIS, CRM, ticketing, accounting, and file storage | Conditional | Good when agents need normalized access to business systems through category APIs. |
| Metorial | metorial.com | Tool and MCP infrastructure for connecting agents to external services | Watch | Track in the agent-tool hub category; evaluate security and connector quality before adoption. |