AGENT OS · SKILLS LIBRARY

Find the right tool. Automatically.

Semantic search across thousands of services. Auto-discovery when something is new.

SKB is the agent's skills library. Instead of pre-configuring every possible tool, agents describe what they need in plain language and SKB returns the exact capability — across Slack, Stripe, Salesforce, Notion, or a service nobody on your team has heard of yet. When a service is unknown, SKB learns it on demand from its public docs.

Service scale
On-demand
Service discovery
Vector
Capability search

Semantic capability match

// agent intent
"post the Q3 numbers to the team chat"
slack.send_message
domain: chat · score: 0.93
match
teams.post_message
domain: chat · score: 0.87
alt
discord.send
domain: chat · score: 0.71
alt

SKB embeds intent + capabilities and returns the best match by score and domain.

Why hard-coded tool lists break agents

The skills library is where most agent projects start to drift.

✗ Pre-configuration is a tax on every release

Every new tool needs OAuth setup, schemas, prompt examples, and tests. Six tools is a sprint; six hundred is a roadmap.

✗ Hard-coded skills break when infra changes

Rename a column, swap a provider, and a thousand prompts go stale at once. Drift is silent until something blows up in production.

✗ Skills lock you to today's tools

If your agent says 'search Notion', migrating to Confluence is a rewrite. Tool names should not be load-bearing.

What SKB does

Four properties that make the skills library survive change.

Semantic capability search

pgvector + OpenAI embeddings match an intent like 'post to the team chat' to the right send_message capability — across Slack, Teams, Discord, or whatever the org actually uses.

LLM-assisted auto-discovery

When a service isn't yet in the catalog, SKB pulls its docs (including JS-heavy SPAs via the Playwright Renderer), extracts auth + endpoints + capabilities with Claude, and queues it for admin verification.

#

Domain tags survive tool swaps

Playbooks reference 'search document store', not 'search Notion'. Swap Notion for Confluence and the playbook keeps working — the binding happens at execution time against the org's actual SKB account.

Plays well with MCP

MCP servers light up alongside SKB connectors. The community ships breadth, SKB adds the semantic layer, domain tags, and credential isolation on top.

From intent to executed capability

How the skills library is queried in production.

1

Embed

The Orchestrator turns the agent's intent into an embedding.

2

Match

SKB returns the top-k capabilities by score, scoped to domain tags.

3

Discover

If nothing scores high enough, SKB launches discovery against public docs.

4

Verify

Auto-discovered services queue for admin review before they're trusted in agent runs.

Skills library, honestly compared

What SKB gives you on day one vs. the alternatives.

Capability Interactor SKB ChatGPT / Cowork Build from scratch
Find the right tool automatically Yes — semantic Pre-configured only Yes — 2+ months
Unlimited service scale Yes — discovery Plugin marketplace You own the catalog
Survives infra changes Yes — domain tags Breaks on rename Depends on you
MCP compatibility Yes — alongside SKB Limited DIY
Time to first connector Minutes Marketplace-bound Days per service

Run a query through our SKB

Give us a sentence in plain English. We'll show you the capability match, the discovery flow, and the secured execution.