
Redact PII before your model searches the web
Build a loop where Rampart redacts personal details on your machine, Tabstack researches the redacted question, and your own model writes the reply. Four short Python files.
Tabstack blog
Practical guides on building answers from web sources, returning clean page content and typed data, integrating with model stacks, and completing public browser tasks.
Featured

Build a loop where Rampart redacts personal details on your machine, Tabstack researches the redacted question, and your own model writes the reply. Four short Python files.
Topics
How current answers are built, when search is enough, and when the research loop should run outside the model.
Read this topicClean Markdown, schema-matched JSON, recurring source sets, and web-to-pipeline patterns.
Read this topicWorking examples for Python, CLI, Hermes, LangChain, LlamaIndex, Mastra, Vercel AI SDK and other supported paths.
Read this topicWhen browser interaction is required, how hosted Automate works, and when to use Pilo.
Read this topicAll posts

Hosted, big cloud, EU, rented GPUs, hardware enclaves, or local: what each option keeps, who can see your prompts, and what it costs.

A webpage becomes schema-matched JSON by one of three routes: read the structured data it already carries, map selectors to fields, or describe the shape and let an extractor fill it. When each one holds up, when it breaks, and how to validate what comes back.

What you maintain when your application runs its own search, fetch, and citation loop, and the hosted alternatives that return sources or a cited answer instead.

Three routes: a hosted provider's search tool, a search-and-cite pipeline you build, or one Research call that returns the report and its cited sources for any model. What each returns, what you own, and how to check the citations.

When an API has docs but no OpenAPI spec, the usual move is to write the client by hand and bury your guesses in the code. I built a small CLI that turns the docs pages into a contract first, with every gap marked as missing, and then generates the client from that.

Local models can't see anything after their training cutoff, and they can't open a link. I built a single-file Tabstack tool for Open WebUI that gives them cited live research and clean page reads, and the live runs changed the design more than I expected.

A small CLI that asks a model you run the same questions with and without web research, and shows you the answers side by side.

Use seven release gates to check a current-answer feature. Define acceptable answers, test failures, and attach the evidence your team needs to ship.

I asked a 22-line Tabstack Research demo what had changed in the latest Laravel release, and it told me 12.x. Oops. Here's how I took it to production one commit at a time: routing, deadlines, retries, caching, and checking the answers against the changelog.

Define expected answer elements, freeze source evidence, and review a live-web feature with a small evaluation set that keeps failures and missing scores visible.

Compare search APIs, agentic search, Research APIs, and DIY loops. Choose by required output, control, data boundaries, and operational ownership.

What a Research call sends, who processes it, what we retain afterward, and the detailed data collection setting that controls it.

Record a Tabstack Research request, preserve its cited pages, and check the answer against its sources with a Python review sheet.

How to check whether citations in an AI answer support its material claims, scope, and date.

Nine places web research fails after search has already succeeded, each with a signature you can observe, the evidence that identifies it, and the smallest test that catches it.

A second path at the application layer sends current questions to a managed Research call and returns a cited result in your own response shape, with the model left exactly as it is.

A web research system you can trust with a current question does seven separate jobs. It defines what a complete answer looks like, discovers candidate sources, retrieves them, extracts evidence, checks that evidence for gaps, reconciles conflicts, and writes a cited result.

We are comparing two complete ways of answering current public-web questions: a pinned search + fetch + model pipeline the application runs itself, and a single managed Research call. This article publishes the pre-registered method. It does not publish a winner.

Clone the example, set one environment variable, and run one command to send a current question to Tabstack Research. The call streams progress, then writes a Markdown report and a machine-readable list of cited pages.

A web search API can return useful titles, URLs, and excerpts. A cited answer requires a research loop after search: define the question, select and read sources, identify gaps, reconcile evidence, write the answer, and preserve the claim-to-source connection.

Install @tabstack/llamaindex, pass the tools to an agent, and let it read pages, pull typed fields, and research questions across sources on demand.

A step by step tutorial: install @tabstack/openai-agents, hand the tools to an Agent, and understand why the adapter registers them with strict mode off.

Install @tabstack/claude-agent, wire its ready-built in-process MCP server into a query() call, and give Claude extraction, research, and browser tools.

A step by step tutorial: install @tabstack/mastra, spread the tools into an Agent, and give it schema-enforced extraction, research, and browser automation.

A step by step tutorial: install @tabstack/eve, add one-line tool files, and give your agent schema-enforced extraction, research, and browser automation.

A step by step tutorial: install @tabstack/ai, give the model web tools, and stream a cited answer token by token instead of making users wait.

A step by step TypeScript tutorial: install @tabstack/langchain, call tools directly, then assemble an agent that reads a competitor's pages and produces a structured brief.

A step by step tutorial: install langchain-tabstack, call a tool directly, then wrap the whole toolset in a LangChain agent that reads the web and shows its working.

Most web-data jobs are just fetch-and-clean, not an agent. Turn any URL into clean markdown or typed JSON, cheaply. Why it's the right default.

Provision Tabstack API keys natively via the Stripe CLI. Programmatic billing, automated account lifecycle management, and browsing infrastructure built for autonomous AI agents.

Generate a structured JSON schema from any URL with Schema Source by Tabstack, then hand it to Extract for matching JSON with no manual schema work.

A technical look at the prompt injection vulnerability in Tabstack and the code changes made to fix it.

Stop your web agents from guessing. Pilo Interactive Mode brings human-in-the-loop capabilities to browser automation, allowing agents to pause and ask for the exact context they need to complete complex tasks.

Tabstack's new pricing is here! Scale your workflow without interruption using our new free Trial, Individual Pay As You Go plan, or predictable monthly subscriptions.

The new effort parameter lets you control speed vs. rendering depth when fetching content with Tabstack API. Start fast, escalate when needed.

Meet Pilo: an open source agentic web automation engine by Tabstack. Build reliable, self-healing AI web agents that truly understand the modern web.

Offload browser orchestration and agentic loops to Tabstack. Bridge the synthesis gap with a research primitive that delivers verified, cited web data.

Tabstack is the web execution layer for AI agents. Automate workflows and turn web content into clean data with a developer API backed by Mozilla.
START FREE
Start with 10,000 free credits. No credit card required.
curl -fsSL https://tabstack.ai/install.sh | sh