The internet was built for humans. Every CAPTCHA, every cookie banner, every login wall — all of it assumes a person with a mouse and some patience. We built RebaseKit to change that. One deployment. One gateway. Every primitive an AI agent needs to actually do things on the web.
Here’s where we are, and where we’re going.
What We’ve Shipped
RebaseKit is a monorepo of 17 services, all running behind a single nginx gateway at api.rebaselabs.online. No 17 separate deployments. No 17 separate subscriptions to manage. One stack.
The APIs
PII API — Detect, mask, redact, and synthesize personally identifiable information. Handles text, files, HTML, JSON, and CSV. Includes an encrypted vault so agents can store and retrieve real values behind opaque tokens. Batch processing, streaming, custom entity types, full audit trail.
WebTask API — The core of everything agents need to interact with the live web. Scrape pages (plain HTTP or full JS rendering), extract structured data, run AI-powered schema extraction, research tasks that chain multiple steps, take screenshots, map sitemaps, consume RSS feeds, and run async jobs that survive connection drops.
LLM Router API — A unified interface to Anthropic, OpenAI, Groq, and Mistral. One endpoint, multiple providers. Includes vision (image inputs), tool-calling, streaming, conversation threads with persistence, reusable prompt templates, batch jobs, token usage billing per API key, and side-by-side model comparison. Embeddings too.
Document Parse API — Give it a PDF, DOCX, XLSX, PPTX, CSV, TXT, .eml, or Markdown file. Get back structured text and metadata. Built so agents can ingest documents without caring about format.
Entity Enrichment API — Company and person intelligence. Enrich an email or domain into a full profile: industry, size, tech stack, social links, news mentions, people discovery, ICP scoring. The kind of data that used to require a $500/month data enrichment subscription.
Agent Proxy API — A managed HTTP proxy with profile management, persistent cookie sessions, and snapshot diffing. Agents can log in once, hold a session, and detect when a page changes between visits.
Agent Storage API — A key-value and file store purpose-built for agents. List, search, copy, move, batch-delete, patch metadata, set TTLs, import from URLs. Namespaced under prefixes like sessions/, memory/, artifacts/, exports/. Agents need to remember things. This is where.
Agent Email API — Full email I/O. Parse inbound messages, thread conversations, extract structured data from email body, send via SMTP, AI-draft replies, and manage an IMAP inbox (check, mark read, move between folders, delete). Agents can have an actual email inbox.
Code Exec API — Sandboxed Python and JavaScript execution. Give an agent a code interpreter that can’t escape the sandbox.
Screenshot API — Playwright-backed screenshots and PDF generation. Full JS rendering, mobile viewports, custom wait conditions. If a human can see it in a browser, an agent can capture it.
PII API (Vault) — Already covered above, but worth calling out separately: the vault mode lets agents redact PII before sending data to third-party LLMs, then re-hydrate results. A real privacy layer for agentic workflows.
Email Builder API — Render, validate, spam-check, and assess send-readiness for HTML email templates. Agents composing emails can verify them before they hit a real inbox.
Email Validator API — DNS MX lookup, SMTP handshake, disposable domain detection. Before an agent sends anything, it can know whether the address is real.
Data Transform API — Convert between JSON, CSV, TSV, YAML, XML, and XLSX. Streaming ndjson. Natural language query execution over structured data. Fuzzy deduplication. The boring plumbing that makes everything else work.
Web Intel API — Deep web extraction: structured schema.org data, tech stack fingerprinting, sitemap crawling, link graphs, feed discovery, raw markdown. The intelligence layer on top of raw HTML.
Mock LLM API — A full LLM API mock that behaves like OpenAI’s interface. Scripted scenarios, deterministic embeddings, custom response sequences. For testing agent workflows without burning real tokens.
Mock Any API — Mock any JSON API. Define a response schema, get a consistent fake endpoint back. For building and testing integrations before the real service is ready.
Audio to Text API — Whisper-backed transcription with language detection, timestamps, SRT/VTT export, and batch processing. Agents that need to understand audio.
The Architecture
One Coolify deployment. One docker-compose.yml. One nginx gateway routing 17 services by URL prefix. Every service has its own auth layer (API keys, tiers, usage billing), its own SQLite backing store, and its own versioned Docker image.
The pattern: api.rebaselabs.online/{service}/api/{endpoint}
nginx strips the prefix. The service sees a clean path. Swagger docs work correctly through the gateway. We do not create new repos per API. We do not create new Coolify deployments per API. One repo, one stack, one domain — and it scales horizontally.
What’s Next
We’re past the “prove it works” phase. Here’s what we’re building toward:
Agent Memory That Actually Works — Right now agent-storage-api is a fast key-value store. The next version adds SQLite FTS5 full-text search over keys and metadata, so agents can surface relevant past context without knowing exact keys. The building block for real agent memory.
Smarter LLM Routing — Prompt caching (Anthropic), context window management, automatic compression of long threads. LLM costs compound fast in agentic workflows. The router needs to fight that aggressively.
Multi-Step Web Actions — WebTask today is largely read-only: scrape, extract, research. The next version adds multi-step form filling and cookie/session management. Agents that can act on pages, not just observe them.
The Bigger Picture — Every API here is a tool. Tools become useful when an agent can compose them. The vision is an agent that can: receive an email → enrich the sender → research their company → draft a reply → validate the address → send it — all as a single orchestrated workflow, no human in the loop. We’re building the primitives to make that work at production scale, with real billing, real auth, and real privacy controls.
The internet has seventeen years of infrastructure built for browsers. We’re building the eighteen months of infrastructure that makes it work for agents. We’re not done. But the foundation is real.
More soon.