Methodology
As of 2026-11-21, AgentSidekick holds 73,820 active products (9,469 with a profile page, 259 checked by a researcher), collected from 12 connected sources and rules. Each fact on the site carries a mark for the kind of evidence behind it, derived data never overwrites a publisher's own statement, and no compatibility pairing or security property has been tested. This page sets out every source, rule, weight and limit, with the numbers taken from this build's database.
- SRC stated by a publisher or registry
- CUR checked by a researcher
- DER computed by rule
- HEU an estimate
- UNK not stated anywhere
Sources
Read from the sources table of this build's database.
12 sources are enabled. 68,059 of the 73,820 active products (92.2%) come first from the official MCP Registry; the rest come from GitHub, Open VSX, skill and plugin files, and curated research. The terms column is our reading of each source's conditions, recorded when it was connected.
| Source | Kind | Terms and what the data means | Products first seen here | Last run |
|---|---|---|---|---|
| Claude plugin marketplace manifests (external)claude-plugin-marketplaces | repo-files | Reads .claude-plugin/marketplace.json from public GitHub repos. | 742 | ok · 2026-11-21 |
| GitHub Advisory Database (external)github-advisories | api | Advisories GitHub has reviewed for npm, PyPI, NuGet and crates.io packages only; no advisory found is not evidence a package is safe, and OCI/mcpb packages are not covered. | — | ok · 2026-11-21 |
| GitHub issue search counts (external)github-issue-search | api | Counts of public issues/PRs matching a phrase. A demand proxy, not a market-size measure. | — | ok · 2026-11-21 |
| GitHub repository metadata (external)github | api | GitHub GraphQL API. Stars, forks, issues, releases, licence and push dates are read as-is. | — | ok · 2026-11-21 |
| GitHub topic search (external)github-topics | api | GitHub REST API. Discovery is by self-applied repository topics, so product type is auto-classified and marked "auto-discovered". | 3,291 | ok · 2026-11-21 |
| Official MCP Registry (external)mcp-registry | registry-api | Public REST API (v0). Namespaces are ownership-verified by the registry (GitHub login or DNS). | 68,059 | ok · 2026-11-21 |
| Open VSX Registry (external)openvsx | registry-api | Public API operated by the Eclipse Foundation. Download counts are Open VSX downloads only (not VS Code Marketplace installs). | 148 | ok · 2026-11-21 |
| SKILL.md files in public GitHub repos (external)github-skills | repo-files | Reads the name/description frontmatter of SKILL.md files. Each listing links to the file and inherits the repository licence. | 1,321 | ok · 2026-11-21 |
| npm and PyPI download statistics (external)package-registries | api | Public registry download counts for the last 30 days; they include CI runs, mirrors and bots, so they measure package fetches, not users. | — | partial · 2026-11-21 |
| Curated researchatlas-curated | curated | Researched against official vendor sites, docs and repositories. Every record lists the URLs checked and the date. | 259 | ok · 2026-11-21 |
| Rule-based classifieratlas-classifier | derived | Keyword rules from data/taxonomy/taxonomy.json. Derived, not a source fact. | — | ok · 2026-11-21 |
| Scoring engineatlas-intelligence | derived | Deterministic scores computed from the tables above. See /methodology/. | — | ok · 2026-11-21 |
Researched but not connected
9 sources were examined and left out. Each one either needs credentials the project does not have, restricts reuse, or has no public interface. Products listed only there are missing from the catalogue.
| Source | Key needed | Why it is not connected |
|---|---|---|
| Cursor Marketplace (external) | No | No public API found. |
| Glama MCP directory (external) | Yes | API key required; licence demands visible Glama attribution and a link to each Glama listing. |
| OpenAI GPT Store / Apps directory (external) | No | No public listing API. |
| Product Hunt (external) | Yes | API token required; non-commercial use by default. |
| PulseMCP (external) | Yes | v0.1 API is partner-only (API key + tenant ID by email). |
| Smithery (external) | No | API shut down 2026-10-31 per smithery.ai/future. |
| VS Code Marketplace (external) | No | Terms restrict use to Microsoft products; Open VSX is used instead. |
| mcp.so (external) | No | No public API; robots.txt disallows /api/. |
| skills.sh (external) | Yes | API accepts only a Vercel OIDC token. |
Pipeline
pipeline/run.mjs, in the order it runs. Each step is isolated: a failure is logged and the next step still runs.
A zero-dependency Node pipeline fills a SQLite database; the site is then built from that database with no network access. Raw payloads are stored when they change, every step writes a row to ingestion_runs with counts and status, and errors go to ingestion_errors. Pages never write to the database.
- 01Official MCP Registry ok · 2026-11-21
Every server entry from the public v0 API: name, description, versions, packages, remote endpoints, environment variables.
- 02GitHub topic search ok · 2026-11-21
Repositories tagged with agent-related topics. Product type is inferred from the topics, so these records are marked auto-discovered.
- 03Curated seed records ok · 2026-11-21
Platforms, protocols, frameworks and infrastructure researched by hand from data/seed/*.json, each with source URLs and a check date.
- 04GitHub metrics (first pass) ok · 2026-11-21
Stars, forks, open issues, licence, push and creation dates for linked repositories.
- 05SKILL.md files ok · 2026-11-21
Name and description frontmatter of SKILL.md files in public repositories, plus a static text inspection of each skill.
- 06Plugin marketplaces ok · 2026-11-21
.claude-plugin/marketplace.json manifests in public repositories.
- 07Open VSX ok · 2026-11-21
Agent-related editor extensions and their Open VSX download counts.
- 08GitHub metrics (second pass) ok · 2026-11-21
Metrics for repositories added by the skill, plugin and Open VSX steps.
- 09Classify ok · 2026-11-21
Applies data/overrides.json, then keyword rules: taxonomy, integrations, industries, platform mentions, registry-metadata signals, completeness and notability scores, duplicate candidates.
- 10Advisories ok · 2026-11-21
GitHub Advisory Database entries for npm, PyPI, NuGet and crates.io packages of profiled products.
- 11Demand ok · 2026-11-21
Counts of public GitHub issues containing "{integration} mcp", one search per integration.
- 12Intelligence ok · 2026-11-21
Segment statistics and labels, the competitor graph, and opportunity scores with their evidence.
- 13Package downloads partial · 2026-11-21
npm and PyPI 30-day download counts. Runs last because registry rate limits make it the slowest step.
Steps 1–8 are ingestion (npm run pipeline); steps 9–13 are analysis and can be re-run alone (npm run analyze) after editing the taxonomy or config. The site is regenerated from the result on every build, so reports and answers always reflect the latest run.
Claim kinds
Every field written to a product records where it came from. The database holds 986,701 field records marked source fact, 3,182 curated and 0 AI-generated. On the site, the five marks mean:
- SRC Source fact
- Stated by the publisher or a registry (the MCP Registry, GitHub, Open VSX, a SKILL.md file). Recorded as given and not checked by us. A publisher's description of its own product is a source fact about what it says, not about what it does.
- CUR Curated
- Researched by a person against official sources. The record lists the URLs checked and the date. 259 products are curated.
- DER Derived
- Computed by a fixed, published rule from source data: counts, shares, categories assigned by keyword, segment labels, scores.
- HEU Heuristic
- A rule of thumb with no measurement behind it, such as the monetisation component of an opportunity score. Useful for sorting, not a fact.
- UNK Unknown
- No source we track states it. Shown as "Unknown" or "—" with the reason, never left blank and never guessed.
Precedence rule. When two sources set the same field, the lower-ranked one never overwrites the higher. The rank is enforced in the pipeline's write function (upsertProduct), so derived or AI output can never replace a source fact, and a source update can never replace a curated value:
- Correction override — A maintainer edit in data/overrides.json. Outranks everything.
- Curated curated — Researched against official sources, with the URLs and date recorded.
- Source fact source_fact — Stated by the publisher or a registry, stored as given.
- Derived derived — Computed by a fixed rule from the rows above.
- AI-generated ai_generated — Model output. Lowest rank; none exists yet.
Compatibility
Compatibility is a status, never a yes or no. No pairing has been independently verified: AgentSidekick has not run a test harness. Of the 445,043 combinations of 9,469 profiled products and 47 tracked platforms, the counts by status are below, computed with the same function every page uses.
| Status | Meaning | Pairs | Share |
|---|---|---|---|
| Independently verified | Tested by AgentSidekick with a test harness. No harness has been run, so no pairing has this status. | 0 | 0.0% |
| Officially documented | The platform's or the product's official documentation states this specific pairing. Reserved for curated per-pair records; none are recorded yet. | 0 | 0.0% |
| Vendor claimed | The publisher's own description names the platform. A claim by the vendor, matched by text, not tested. | 2,268 | 0.5% |
| Protocol-level | The platform officially documents support for the standard the product implements (MCP, the Agent Skills format, Claude plugins). This pairing was not tested. | 227,073 | 51.0% |
| Incompatible | The platform is documented as not supporting that standard. | 0 | 0.0% |
| Platform discontinued | The platform is discontinued or winding down; its own notice is shown. | 7,443 | 1.7% |
| Unknown | No evidence either way, or no compatibility evidence is collected for this product type. | 208,259 | 46.8% |
Platform support comes from curated platform records: 37 of 47 platforms document MCP support and 27 document the Agent Skills format. A vendor claim comes from matching a platform's name or aliases (data/config/platform-aliases.json, with exclusions for ambiguous words) in the publisher's own name and description; 4,528 such mentions are on record. A "community-reported" status is reserved for user submissions, which need a backend that does not exist.
Classification
Categories, capabilities, integrations and industries are assigned by keyword rules, so every tag is derived. Of the 75,235 unmerged products, 35,291 (46.9%) match at least one taxonomy keyword; the other 39,944 have no category. 6,595 (8.8%) match an integration.
- Taxonomy. 454 active nodes in four levels (ecosystem 9, category 44, subcategory 96, capability 305) with 2,482 keywords, in
data/taxonomy/taxonomy.json. A product matched to a node also inherits every ancestor; inherited tags are marked as such. - Matching. Whole-word, case-insensitive, against the product name, description and repository topics. The matched word is stored and shown on the product page.
- Integrations and industries. 111 integrations matched by their aliases, 28 industries by keywords. A product is never tagged with its own name as an integration, and platforms are not tagged as integrations of each other.
- Stoplist. 17 keywords are ignored because they name a platform or protocol rather than a capability:
agent skills,claude code,claude code plugin,claude skills,cline,codex cli,coding agent,copilot,cursor ide,fastmcp,gemini cli,mcp client,model context protocol,opencode,remote mcp,skill.md,windsurf. Platform names are captured separately as vendor claims. - Product type is a source fact for registry entries (MCP servers by definition), skill files and plugin manifests, and is inferred from repository topics for GitHub discoveries.
Keyword rules miss products that do not name what they do, and generic words can produce false matches. Editing the taxonomy file and re-running analysis re-classifies the whole catalogue without code changes.
Evaluation
Product pages show 8 separate dimensions instead of a single score; there is no aggregate, because a total would hide which evidence is missing. Each has a level (good, middle, low or unknown), a status word and the evidence sentence.
| Dimension | Rule |
|---|---|
| Source authority | Good when the record was curated; middle when published to the official MCP Registry by the namespace owner; low when auto-discovered from repository metadata. |
| Publisher verification | Good when the MCP Registry verified the namespace (GitHub login or DNS); middle for a GitHub organisation or named vendor; otherwise not verified. Confirms who published, not what it does. |
| Maintenance activity | From the last push to the repository: active within 90 days, slowing within 365, inactive after that; archived repositories are low. Unknown without a public repository. |
| Licence clarity | Good when a licence identifier is on record; low when a repository exists but GitHub detected no standard licence file. |
| Installation clarity | Good when packages or remote endpoints are declared in registry metadata; middle when only an install or docs link exists. |
| Documentation | Good when a documentation URL is on record; middle for repository or website only. Documentation quality itself is not assessed. |
| Compatibility evidence | Middle at best: vendor-claimed when the publisher names platforms, protocol-level for MCP servers, skills and plugins. Never good, because nothing is tested. |
| Security review | Always "Not reviewed". AgentSidekick has audited no product; automated signals are listed separately. |
Signals and advisories
Signals describe what a product references. They are not a safety verdict, nothing is executed to produce them, and the absence of a signal is not evidence of safety. Every product's security review reads "Not reviewed".
| Signal | Method | Raised when | Products |
|---|---|---|---|
| remote-endpoint | Registry metadata | The publisher declares one or more remote URLs in registry metadata. | 47,190 |
| local-process | Registry metadata | The publisher declares a package that runs locally over stdio (npm, PyPI, OCI…). | 26,176 |
| requires-secret | Registry metadata | A declared environment variable is marked secret (an API key or token). | 10,204 |
| shell-execution | Static text inspection | The SKILL.md instructions contain a shell code block or an install command (npx, pip install, chmod +x…). | 634 |
| bundled-scripts | Static text inspection | The skill folder contains script files (.sh, .py, .js, .ts, .rb, .ps1). | 297 |
| credential-access | Static text inspection | The instructions mention an API_KEY / TOKEN / SECRET variable, "api key" or a .env file. | 169 |
| network-access | Static text inspection | The instructions mention curl, wget, fetch(), requests.get/post or an api. host. | 160 |
| destructive-commands | Static text inspection | The instructions mention rm -rf, git push --force, git reset --hard or DROP TABLE. | 26 |
| external-scripts | Static text inspection | The instructions pipe a downloaded script into a shell (curl … | sh). | 9 |
Static text inspection reads a skill's SKILL.md and its file list with regular expressions. A skill that downloads and runs code without naming it in its text raises nothing, and a skill that documents rm -rf as something to avoid raises a signal.
Advisories come from the GitHub Advisory Database for the npm, PyPI, NuGet and crates.io packages of 2,940 profiled products. 111 advisories are recorded against 48 products. AgentSidekick stores GitHub's vulnerable range and first patched version as given and does not decide whether the listed version is affected. OCI and MCPB packages are not covered, so "no advisory" means none was found in that database.
Segments
Calculated 2026-11-21.
A segment is a taxonomy node, an integration or an industry. For each of the 593 segments the pipeline counts products, launches in the last 90 days and the 90 before (by publish or repository creation date), total and median GitHub stars, the top project's share of stars and the Herfindahl index of star concentration.
Relative momentum. A growing ecosystem makes every segment look like it is accelerating, so momentum is measured against the whole catalogue:
momentum = round(50 + 50 · tanh(ln(((new + 1) / (prev + 1)) / base)))
base = (catalogue launches, last 90 days + 1) / (catalogue launches, previous 90 days + 1)
50 means the segment grew exactly as fast as the catalogue. With fewer than 5 launches across both windows the score is not computed ("insufficient").
Labels. Size is the percentile of the segment's product count among its peers (nodes at the same level, or all integrations, or all industries). At the 80th percentile or above a segment is Crowded, qualified as outpacing the catalogue at momentum ≥ 62 or slowing at ≤ 38. At the 30th or below it is Emerging with momentum ≥ 62, otherwise Thin. In between: Outpacing the catalogue, Lagging the catalogue or Established. A segment with no products is Empty.
| Label | Segments | Share |
|---|---|---|
| Thin | 157 | 26.5% |
| Lagging the catalogue | 155 | 26.1% |
| Established | 83 | 14.0% |
| Crowded, slowing | 64 | 10.8% |
| Outpacing the catalogue | 53 | 8.9% |
| Crowded | 34 | 5.7% |
| Crowded, outpacing the catalogue | 25 | 4.2% |
| Emerging | 20 | 3.4% |
| Empty | 2 | 0.3% |
Opportunity score
188 opportunities are scored: 108 integration gaps (one per integration with a demand measurement; 111 measured) and 80 coverage-gap hypotheses (capabilities whose parent subcategory holds at least 20 products; the 80 with the highest overall score are kept). Each component is 0–100; the overall score is the weighted sum, rounded to one decimal. Every component's evidence is stored and shown, marked measured or heuristic.
| Component | Weight | Integration gap | Coverage-gap hypothesis |
|---|---|---|---|
| Demand | 0.25 | Percentile of public GitHub issues containing "{integration} mcp", among all integrations measured. | Percentile of the parent subcategory's combined GitHub stars, multiplied by 0.6 because it is a proxy. |
| Competition gap | 0.20 | 100 minus the percentile of products that are complete (quality ≥ 0.6) and pushed in the last 180 days. | 100 minus the percentile of the capability's product count among all capabilities. |
| Monetisation | 0.20 | 35, or 65 for an enterprise system; +20 for a group businesses commonly pay for (CRM, accounting, payments…). Heuristic. | 65 under business, commerce, finance, infrastructure or data ecosystems, otherwise 40. Heuristic. |
| Growth | 0.15 | The segment's relative momentum score (below); 50 when there are too few launches to measure. | Same, for the parent subcategory. |
| Feasibility | 0.10 | 80 when the integration has a public API on record, otherwise 35. Heuristic. | A neutral 60: not assessed. |
| Distribution | 0.10 | Share of tracked platforms documenting MCP support (37 of 47), as 0–100. | Same. |
Proxy discount. Demand for a single capability is not measured; it is inferred from the parent subcategory and multiplied by 0.6 (a 40% discount) before it enters the score. Coverage-gap hypotheses therefore carry "Low" confidence. An integration gap is "Medium" confidence only with at least 50 issue mentions and 10 products.
A high score says the measured inputs line up, not that a market exists: issue mentions measure developer discussion, not buyers, and a low product count can mean low demand.
How rankings work
Every ranked list states its rule in one sentence on the page. Most rank by GitHub stars of the product's own repository; a product inside a shared monorepo has no stars of its own and is listed after. Stars measure attention on a repository, not installs, quality or safety.
Where no single metric fits, lists use notability, a derived sort key:
notability = log10(stars + 1) [× 0.35 for a subfolder of a shared repo]
+ 0.6 · log10(Open VSX downloads + 1)
+ 3 if curated
+ 0.4 if published by an organisation
No paid influence. There are no sponsored placements. Nobody can pay to change a ranking, an evaluation or a compatibility status. If placements are ever sold, they will be labelled and kept out of rankings and evaluations (see pricing).
Which pages are indexable
Thin pages are built but kept out of search engines (noindex), with the reason recorded. The gates live in src/lib/seo.ts and the pipeline's completeness score.
- Profile pages exist only for products with data completeness of at least 0.6, or a curated record (9,469 of 73,820). Completeness adds 0.25 for a description of 60+ characters (0.1 for 25+), 0.2 for a repository or website, 0.2 for a keyword category, 0.15 for install or docs information, 0.1 for a licence or pricing model and 0.1 for an integration or platform mention.
- A profile is indexable when curated, or when it has a description of at least 80 characters, at least one capability or integration, and completeness of at least 0.75. 6,358 of 9,469 profiles pass (67.1%).
- Listing pages (categories, integrations, platforms, companies…) need at least 5 products to be indexable.
- Answer pages are published only when their query returns at least 5 qualifying products or a complete comparison: 26 of 26 pass.
No FAQPage or HowTo markup is used: Google no longer shows either. Pages carry Article, Dataset, SoftwareApplication (without ratings we do not have) and breadcrumb markup only.
Known limits
- One metrics snapshot. Stars and downloads were captured on 1 day, so growth is measured from launch dates, not from changes in stars.
- Keyword misses. 39,944 products (53.1%) match no category; others are mis-tagged by generic words.
- Registry self-description. 92.2% of products come from the MCP Registry, where every field is the publisher's own claim.
- No pricing data. No pricing plan has been recorded for any product; pricing is shown as Unknown rather than estimated.
- No LLM enrichment. No model has been run over the catalogue; there are no AI-generated fields.
- Counts are of listings, not usage. A product counted here is a record in a registry or repository; it says nothing about installs, users or revenue. Download counts include CI runs and mirrors.
- Coverage gaps. The 9 sources above are not connected, so products listed only in vendor marketplaces (Cursor, VS Code, the GPT Store) are under-represented.
- Nothing tested. No compatibility pairing and no product's security has been tested by AgentSidekick.
Corrections
There is no submission form yet: that needs a backend. Until one exists, corrections are applied by maintainers in data/overrides.json, which outranks every source and is re-applied on every analysis run. It currently holds 168 merges of duplicate records, 56 field corrections, 1 type corrections and 3 integrations excluded from demand measurement because their names are too ambiguous to search.
To report an error, send the page URL, the field that is wrong and a source that shows the correct value: Contact address not configured.
Data licence
Facts taken from public registries and repositories (names, descriptions, versions, stars, licences) belong to their publishers and are reproduced under each source's terms above. Curated records, classifications, scores and other derived data are © AgentSidekick. A licence for reuse of the dataset has not been chosen yet; until it is, no licence is granted beyond what each source's own terms allow.
The data is published as JSON: products, platforms, taxonomy, integrations, segments, opportunities, compatibility matrix and changes. Each file states its claim kind and generation date.