Small, sharp, and built to compose later

None of these need a paragraph of pitch. Each one does a single job an agent already asks for, and most of them feed straight into the composite packs shipping alongside them today.

AgentOps: watching agents watch themselves

agent-trace-brief takes a raw execution trace (a JSON array of steps and tool calls, or a plain log) and returns an ordered summary, the failure points, and a root-cause hypothesis grounded only in what the trace shows. mcp-tool-risk-score reads an MCP server's tool manifest and flags which tools can delete, spend, or exfiltrate before you grant them. tool-call-diff compares two tool-call sequences and tells you exactly what changed between runs. eval-case-generate drafts {input, expected, rubric} regression cases from a described task, so a production incident turns into test coverage instead of a war story.

Browser workflow: reading the DOM before you break on it

dom-change-diff compares before/after HTML and tells you which of your selectors still resolve, which moved, and which vanished. form-fill-plan reads a form's HTML plus a stated goal and returns a field-by-field fill plan, so an automation script doesn't have to guess which input is the email field. browser-flow-digest takes a recorded action trajectory and turns it into a reusable, readable summary of the workflow.

Model routing: pick the right model without guessing

llm-cost-estimate prices a token count across a model list, in one call, no spreadsheet. model-route-recommend takes a task description and constraints (latency budget, cost ceiling) and recommends which model tier fits.

Retail and procurement: policy text into structured data

return-policy-extract reads a merchant's return policy text and pulls out the window, restocking fees, and exceptions as structured fields. warranty-parse does the same for warranty terms: coverage length, exclusions, claim process. sku-normalize takes a messy product title and category and returns a normalized SKU-shaped record.

Learning: syllabus and paper into study material

syllabus-parse turns a course syllabus into a structured week-by-week breakdown with readings and due dates. paper-to-flashcards reads a research abstract or excerpt and generates flashcards from it. reading-plan-generate takes a topic, a skill level, and a time budget and returns a reading sequence.

Support ops: tickets in, structure out

ticket-cluster groups a batch of support tickets into clusters by underlying issue. sla-extract reads a contract or SLA text and pulls out response times, resolution windows, and credit terms. escalation-brief takes a support thread and account context and drafts an escalation summary. duplicate-ticket-detect checks one new ticket against a candidate list and flags likely duplicates.

Civic and science lookups

grant-lookup searches by keyword across open grant listings. civic-meeting-brief reads civic meeting text (minutes, transcripts) and summarizes the decisions and votes. ncbi-taxonomy-lookup resolves an organism name to its NCBI taxonomy record.

Wiki and search primitives

wikipedia-search, wikidata-resolve, and wikipedia-pageviews round out the wiki-lookup set that several composites (entity-knowledge-card among them) already call underneath.

Why 25 at once

Most of these are one-call answers to questions agents already have to answer themselves by hand, usually by scraping a page or writing a regex. They're priced as smalls because that's what they are. The composites shipping today lean on a handful of them directly, which is the other reason they went in together.