The Problem: Agent Tools Are Siloed Across Platforms
If you're building AI agents, you've probably hit this wall: you need to find the right capability (like "send an email" or "read a calendar"), but it's scattered across platforms. OpenClaw has its own skill registry. MCP has its servers. LangChain has tools. CrewAI has its ecosystem. Composio maintains another set. And so on.
Each platform has its own format, its own documentation style, its own submission process. No one index exists. No single search that says "show me every way an agent can send email across all platforms."
So we built one.
What is AgentSift?
AgentSift.ai is a unified capability discovery engine for AI agents. It indexes 341+ capabilities across 7 major platforms:
- OpenClaw (custom agent framework)
- Model Context Protocol (MCP) (Anthropic's server standard)
- LangChain (tools ecosystem)
- CrewAI (agent orchestration)
- Composio (integration platform)
- OpenAI (function calling)
- NEAR AI (blockchain agent platform)
One search box. Every capability. Fuzzy intent matching so "I want to send emails" finds every email-related capability across all platforms, not just exact string matches.
The Safety Angle: Informed Consent for AI Agents
Here's the part most capability directories skip: safety ratings.
Every capability on AgentSift gets:
- Permission level (read-only, write, destructive, network)
- Risk rating (low, medium, high)
- Detailed permission breakdown (what data it accesses, what actions it takes)
Why? Because when you're giving an agent the ability to "manage files," you need to know if that means "read filenames" or "recursively delete everything." No more surprises. Informed consent for agents, just like informed consent for users.
The Tech Stack
We built this in about 6 hours during a weekend sprint. The stack:
- Next.js 15 (App Router) — fast, server-rendered capability search
- Fuse.js — fuzzy search for intent matching
- PostgreSQL — structured capability data
- Nightly data refresh — automated scraping and syncing from each platform
The search algorithm weighs:
- Intent keywords (fuzzy match on description, not just exact titles)
- Platform tags (filter by MCP, LangChain, etc.)
- Risk level (show only safe capabilities, or include everything)
How We Sourced the Data
Each platform has its own capability format:
- MCP publishes JSON schemas
- LangChain has tool class definitions
- OpenClaw uses YAML skill manifests
- Composio maintains an API registry
We wrote platform-specific scrapers that:
- Fetch the latest capability definitions
- Normalize them into a unified schema
- Auto-classify permissions and risks based on declared functionality
- Index them for search
The scrapers run nightly. When a platform adds a new tool, it shows up in AgentSift within 24 hours.
Why This Matters
Right now, every agent builder is solving the same problem: capability discovery. They're manually checking multiple registries, cross-referencing documentation, and hoping they didn't miss the perfect tool that's buried three pages deep in someone's GitHub repo.
AgentSift makes capability discovery one search across all platforms. It's like having a universal package manager for agent tools.
And the safety ratings? That's the difference between agents that are useful and agents that are trustworthy.
What's Next
We're adding:
- Submission flow (let developers add new capabilities directly)
- API access (so agents can query AgentSift programmatically)
- Usage analytics (which capabilities are most searched, most integrated)
- More platforms (AutoGPT, LangGraph, custom frameworks)
But the core is live. 341 capabilities. 7 platforms. One search.
Try it: https://agentsift.ai
Built by: Joe & Microwave (@joespano_ @onlyamicrowave)
Stack: Next.js 15, Fuse.js, PostgreSQL
Source: Platform-specific capability scrapers, updated nightly
License: Public search, data aggregated from open-source registries
Top comments (0)