Developers
API & MCP
The same services power the website, the REST API and the MCP server. There is no separate index, so what your agent sees is what the catalog contains.
MCP endpoint
https://biowise.cc/api/mcpStreamable HTTP transport, stateless, read-only. Anonymous access works at a reduced quota; an API key raises it to the key owner's plan.
Your API key
Connect from Claude Code
claude mcp add --transport http biowise https://biowise.cc/api/mcp
Add --header "Authorization: Bearer bw_live_…" to use a key. Verify with claude mcp list.
Connect from any MCP client
{
"mcpServers": {
"biowise": {
"type": "http",
"url": "https://biowise.cc/api/mcp",
"headers": {
"Authorization": "Bearer bw_live_your_key_here"
}
}
}
}Any client that supports the Streamable HTTP transport can use this configuration shape. For clients that only support stdio, the repository ships a small bridge in mcp/: run npm run mcp:stdio and point the client at it.
Other clients
Support for remote MCP servers varies by product and changes often. Rather than claim compatibility this project has not verified, check your client's own documentation for whether it accepts a remote HTTP MCP server, and use the stdio bridge if it does not.Tools
search_repositoriesread-onlySearch the BioWise repository catalog by free text and facets (category, language, framework, licence, maintenance). Returns published public entries with licence, maintenance and provenance.
inputs: query, languages, frameworks, licenses, maintenance, sort, limit
get_repositoryread-onlyFetch one repository by its BioWise slug, including licence verification state, maintenance signals, limitations and any skills generated from it.
inputs: slug
find_alternativesread-onlyFind catalogued alternatives to a repository, ranked by shared categories, topics, frameworks and language. Deterministic, no model call.
inputs: slug, limit
search_skillsread-onlySearch published agent skills by task, stack or target agent. Each result includes compatibility, validation state and a download URL.
inputs: query, agents, languages, repositoryId, validatedOnly, limit
get_skillread-onlyFetch one skill with its latest published version, including the full SKILL.md, supporting file list, licensing position and version history.
inputs: slug, includeContent
get_skill_versionread-onlyFetch an exact, immutable published version of a skill, including its package checksum, so an install can be reproduced.
inputs: slug, version
get_installation_guideread-onlyGet installation instructions for a repository or a skill. States explicitly whether repository steps have been verified by an editor or only extracted from documentation.
inputs: resourceType, slug
search_collectionsread-onlySearch public, curated collections of repositories and skills. Private collections are never returned.
inputs: query, featuredOnly, limit
get_trending_repositoriesread-onlyRepositories with the most recorded movement over 14 days. The response states whether the ranking is measured or a recency fallback.
inputs: limit
match_cluesread-onlySubmit what you read, in a screenshot, a video, a talk or a web page: repository names, owners, URLs, verbatim on-screen text, spoken product names, technologies, and get ranked candidate repositories with the evidence for each. Scoring is deterministic and computed server-side. Read the media yourself and send text; never send the image or the video. Nothing is written, use record_discovery_scan to keep the result.
inputs: repositoryNames, ownerNames, fullNames, urls, domains, technologies, visibleText, interpretation, confidence
Editor tools
These are registered only for a key with the catalog:write scope owned by a moderator or admin. Anyone else does not see them in tools/list. They exist so an editor's own agent session can curate the catalog without BioWise holding an AI provider key: your agent reads the screenshot or the repository, and submits what it concluded as data. Everything they create is a draft, every record is stamped operator-agent with the model label you pass, and candidate scores are recomputed here rather than taken from the caller.
record_discovery_scanwrites draftsPersist an identification so it appears in the admin intake history, whether you read a screenshot, a video or a web page. Re-ranks the clues server-side; the stored candidates are the computed ones, not ones you assert. Optionally marks which candidate was the match. For a video, read it on your own machine and send what you heard and saw, this server never downloads from YouTube, Instagram or anywhere else.
inputs: repositoryNames, ownerNames, fullNames, urls, domains, technologies, visibleText, interpretation, confidence, readBy, medium, sourceUrl, platform, selectedFullName, note
import_repositorywrites draftsImport a repository from its GitHub URL. Fetches metadata from GitHub, detects duplicates, and creates a DRAFT catalog entry, it does not publish. Returns the draft's admin URL.
inputs: url, categorySlugs
submit_repository_analysiswrites draftsAttach catalog metadata you produced yourself (summary, overview, tags, frameworks, use cases, limitations, installation steps, dependencies) to a draft repository. Rejected for published entries so curated copy cannot be overwritten. Stored unreviewed and labelled as agent-written. Installation steps must be copied from the repository's own documentation, do not invent commands.
inputs: slug, repositoryId, readBy, summary, overview, suggestedCategorySlugs, suggestedTags, frameworks, useCases, limitations, installationSteps, dependencies, resourceKind
Where the image goes
Nowhere. Your agent reads the screenshot on your machine and sends only the text it read, names, owners, URLs, visible text. A URL in those clues is parsed for an owner/name pair and never fetched.