Weave Capability Catalog

The catalog is how agents become discoverable and connectable. Every Weave agent publishes a manifest (its ports + capability tags); the catalog defines the shared vocabulary so a Router can match one agent's output to another's input.

Agent manifest

Every agent emits a portable manifest via agent.manifest() (also written into the .weave.zip produced by weaveflow package):

{
  "name": "summarizer",
  "input_schema":  { "type": "text", "shape": null, "format": "markdown", "description": "" },
  "output_schema": { "type": "structured_json", "shape": { "summary": "str" }, "format": null, "description": "" },
  "capability_tags": ["summarization", "english"]
}

Standard data types

Type	Value (Python)	Common use
text	str	prompts, summaries, reports
structured_json	dict	data extraction, structured output
image	str \	bytes	vision, document parsing
code	str	code gen / review / execution
audio	str \	bytes	voice input, transcription
document	str \	bytes	document analysis, drafting
embedding	list[float]	semantic search, RAG
stream	async/sync iterator	live, long-form output

Type compatibility graph

A source type can feed a target type directly (identical) or via an auto-injected transform (compatible). Anything not listed is incompatible and rejected at connect time.

Source → Target	Transform	LLM needed
text → structured_json	extraction prompt → JSON	yes
text → code	retag	no
code → text	retag	no
structured_json → text	json.dumps	no
stream → text	join tokens	no
document → text	passthrough	no

Extend it with register_transform(SOURCE, TARGET, fn) and a row in the compatibility map in weaveflow/types/primitives.py.

Capability-tag vocabulary

Tags are free-form semantic labels, but a shared vocabulary improves match quality. The Router ranks candidates by tag overlap (swap in embedding similarity via score_fn).

Recommended tag families

Family	Example tags
Function	summarization, extraction, classification, translation, generation, review, routing
Domain	web_search, code, legal, medical, finance, support
Modality	vision, audio, speech, document
Language	english, arabic, spanish, multilingual
I/O shape	rag, structured_output, streaming

Conventions

• Lowercase, snake_case, singular where natural (web_search, not WebSearches).
• Prefer a function tag and a domain/modality tag (["summarization", "legal"]).
• Don't encode the LLM or framework in a tag. That's private to the agent.

Reference agents

A minimal starter catalog you can build against (see examples/):

Agent	input → output	Tags
cleaner	text → text	["text", "clean"]
counter	text → structured_json	["text", "stats"]
summarizer	text → text	["summarization"]