HVTracker｜开源AI Agent与MCP服务器信任评分榜单

access：来源确认 · Browse the live trust registry: [hvtracker.net](https://hvtracker.net)

project_type：来源确认 · AI Trust Registry for Open-Source Agents and MCP Servers

# HVTracker

**AI Trust Registry for Open-Source Agents and MCP Servers**

[hvtracker.net](https://hvtracker.net) ranks open-source AI agents and MCP servers by evidence-weighted trust signals, not GitHub hype.

HVTracker is an ever-growing leaderboard that tracks open-source AI agents and MCP servers across multiple categories and publishes public, machine-readable trust data for each project: activity, adoption, transparency, supply-chain safety, identity, provenance, evidence grade, and rank movement.

The core question is simple:

> Which open-source AI agent and MCP-server projects look active, adopted, transparent, and verifiable right now?

---

## Version 4

The production HVTrust score is **runtime-trust calibrated**: the supply-chain base score plus a bounded adjustment for MCP support, external service dependencies, tool/plugin surface, and package-provenance drift (methodology v4.0). A soft ceiling and evidence-first tie-break keep strong agents from piling onto an identical 100 (v4.1), and a v4.2 correction re-derives the score from the base on every build so the bounded adjustment can't compound. v4.3 narrows every runtime signal to declared evidence: dependency markers are read from dependency sections only, plugin surface from shipped plugin paths, and external services from named credentials — a name in a README, a keyword, or an npm script no longer moves a score. Every agent also carries a separate **Evidence Coverage** grade (A–D) showing how many independent signal types back its score. See the [methodology changelog](https://hvtracker.net/methodology/#changelog) and [runtime-trust spec](https://hvtracker.net/spec/runtime-trust/v0.2/).

Accounts and sign-in (GitHub / Google OAuth), watchlist, the side-by-side compare tray, and crawlable static comparison pages at `/compare/<a>-vs-<b>/` shipped in v3.1 and remain live.

---

## What You Can Do

- Browse the live trust registry: [hvtracker.net](https://hvtracker.net)
- Compare agents side by side: [hvtracker.net/compare](https://hvtracker.net/compare/)
- Read category comparison guides: [hvtracker.net/blog](https://hvtracker.net/blog/)
- Use the public API: [hvtracker.net/data/latest.json](https://hvtracker.net/data/latest.json)
- Embed live trust badges in project READMEs

## Repository Docs

- [docs/README.md](docs/README.md) - repo docs index
- [docs/open-core.md](docs/open-core.md) - public/private product boundary
- [docs/AI_HANDOVER.md](docs/AI_HANDOVER.md) - implementation handover for coding agents
- [docs/launch-v1.md](docs/launch-v1.md) - launch and distribution notes

---

## Current Snapshot

- 1,300+ listed projects — open-source AI agents plus MCP servers from the official registry (see [live count](https://hvtracker.net))
- Curated categories spanning MCP servers, coding agents, frameworks, infra, security, and more
- **~30-min** signal refresh cadence (tunable via `SIGNALS_REFRESH_MIN`)
- **~24h** expected full data sweep across sources
- **90-day** per-agent history where available
- JSON feed of agents and articles at [`/feed.json`](https://hvtracker.net/feed.json)
- Railway-hosted site with a small FastAPI edge and generated public pages/data

Newly submitted agents are listed quickly using a pending-only refresh path, then normal cron jobs keep signals fresh.

---

## Why HVTracker Exists

Most AI agent directories are either manual lists or popularity rankings. Stars can tell you what is visible. They do not tell you whether a project has maintainers, a license, package provenance, signed commits, OSSF Scorecard data, or recent activity.

HVTracker combines curation with independently checkable public evidence. The default rank is **HVTrust**, a 0-100 score designed to reward verifiable trust signals and penalize thin evidence.

```text
base = gate(
  confidence x [ Safety(25) + Identity(18) + Transparency(17)
                 + Maintenance(20) + Adoption(20) ]
  - penalties
)

HVTrust = clamp(
  base
  + RuntimeBonuses x min(1, (100 - base) / 20)   # soft ceiling
  - RuntimePenalties,                            # absolute
  0..100
)
```

| Dimension | Max | What it measures |
|---|---:|---|
| Safety / Integrity | 25 | OSSF Scorecard, package provenance, signed commits |
| Identity / Provenance | 18 | Verified listing status and build provenance |
| Transparency | 17 | License and OSSF transparency checks |
| Maintenance | 20 | Freshness and recent commit activity |
| Adoption | 20 | Log-scaled, capped stars and package downloads |

The base score is then runtime-trust calibrated (v4.0+): a bounded adjustment for MCP support, external service dependencies, tool/plugin surface, and package-provenance drift, re-derived from the base on every build. Confidence is based on present vs applicable signal types. Thin evidence limits how high an agent can rank, even if it is popular.

Read the full methodology: [hvtracker.net/methodology](https://hvtracker.net/methodology)

---

## Two grades: Evidence Grade and Evidence Coverage

Each agent carries two independent A–D grades so the score and the depth of evidence behind it are never conflated.

**Evidence Grade** is the trust-score band — how trustworthy the evidence says the project is:

| Grade | Trust score |
|---|---|
| A | ≥ 80 |
| B | ≥ 65 |
| C | ≥ 50 |
| D | < 50 |

**Evidence Coverage** is separate — how many *independent* public signal types back that score (GitHub, package downloads, supply-chain, behavioural, discussion; five possible):

| Coverage | Meaning |
|---|---|
| A | Broad independent signal coverage (≥ 4 of 5 types) |
| B | Strong public evidence with some gaps (3) |
| C | Basic public evidence (2) |
| D | Mostly GitHub-only or thin evidence (1) |

A high score built on thin evidence keeps a low coverage grade, so breadth stays visible. Both are on every agent page and in the public API (`evidence_grade`, `coverage_grade`, `signal_types`).

HVTracker is not a security certification. Missing provenance, Scorecard, or signature data can mean a signal is unavailable, not that a project is unsafe.

---

## Categories

Projects are curated across MCP servers, coding agents, agent frameworks, workflow platforms, browser & computer use, memory & knowledge, research & data, observability & evaluation, security & guardrails, protocols & tool integration, and more. Live per-category counts (they move as the registry grows) are on the site: [hvtracker.net/categories](https://hvtracker.net/categories/).

---

## Public Data API

The public dataset is licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). CORS is open for public endpoints.

| Endpoint | Description |
|---|---|
| [`/data/latest.json`](https://hvtracker.net/data/latest.json) | Current public trust registry snapshot |
| [`/data/agents/{slug}.json`](https://hvtracker.net/data/agents/dify.json) | Per-agent record with history, events, and trust credential |
| [`/data/build_report.json`](https://hvtracker.net/data/build_report.json) | Build integrity report |
| [`/data/signals/scorecard.json`](https://hvtracker.net/data/signals/scorecard.json) | OSSF Scorecard signal cache |
| [`/data/signals/provenance.json`](https://hvtracker.net/data/signals/provenance.json) | Package provenance signal cache |
| [`/feed.json`](https://hvtracker.net/feed.json) | JSON Feed with agents and comparison guides |
| [`/api/v1/mcp/verify?server=<id>`](https://hvtracker.net/api/v1/mcp/verify?server=topoteretes/cognee) | Pre-connect, signed trust verdict for an agent or MCP server. Curated agents return a full verdict; an unlisted **AI repo with ≥1,000 stars** gets a free provisional verdict; others route to `/submit`. |
| [`/api/v1/verify/recent`](https://hvtracker.net/api/v1/verify/recent) | Public feed of the last 100 checked projects (checks are public by default). |
| [`/llms.txt`](https://hvtracker.net/llms.txt) | LLM-readable project summary and key links |

---

## Trust Badges

Listed projects can embed live HVTrust and evidence-grade badges.

```md
[![HVTrust](https://hvtracker.net/badge/<slug>.svg)](https://hvtracker.net/agents/<slug>/)
[![Evidence Grade](https://hvtracker.net/badge/<slug>-grade.svg)](https://hvtracker.net/agents/<slug>/)
```

Example:

```md
[![HVTrust](https://hvtracker.net/badge/dify.svg)](https://hvtracker.net/agents/dify/)
[![Evidence Grade](https://hvtracker.net/badge/dify-grade.svg)](https://hvtracker.net/agents/dify/)
```

The exact snippet is shown on every agent profile page.

---

## Open-Core Boundary

HVTracker uses an open-core model.

- Public: methodology, specs, curated registry entries, current trust scores, public profiles, correction flow, and the public data API.
- Private later: hosted alerts, watchlists, extended history, team workflows, and higher-volume API access.

The code in this repository remains under the repo's existing MIT license. Public registry data remains licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The brand, hosted service, private enrichment, and future enterprise workflows are not implicitly included in that public-data license.

Read [docs/open-core.md](docs/open-core.md) before widening the public API or changing what gets stored internally.

---

## SEO And Comparison Pages

HVTracker publishes crawlable, data-backed comparison pages:

- Category pages: `/categories/<category>/`
- Agent comparison pages: `/compare/<agent-a>-vs-<agent-b>/`
- Blog comparison guides: `/blog/<category>-top-agents/`

These pages are generated from the current registry data, included in `sitemap.xml`, and linked from `feed.json` and `llms.txt`.

---

## How It Works

```text
agents.json ──┐
               ├──> fetch_and_build.py ──> index.html
history/    ───┤                           agents/<slug>/index.html
scorecard-  ───┘                           data/latest.json
cache.json                                 data/agents/<slug>.json
```

1. `fetch_and_build.py` reads curated agents from `agents.json`.
2. Public APIs are fetched in parallel where safe and serially where rate limits require it.
3. HVTrust scores, evidence grades, rank deltas, trust breakdowns, and events are computed.
4. Static pages, JSON endpoints, badges, specs, feed files, sitemap, and build reports are generated.
5. Railway serves the generated site from a persistent volume and refreshes it every 4 hours.

### Build Modes

```bash
python fetch_and_build.py              # full refresh
python fetch_and_build.py --batch 1/6  # one staggered batch
python fetch_and_build.py --pending-only
python fetch_and_build.py --render-only
```

- `--pending-only` refreshes newly listed agents without running a full batch.
- `--render-only` rebuilds pages from cached render state without API calls.

### Manual Production Deploy

Merging is not deploying. Production only changes when someone uploads a
build, and that upload always comes from a **clean worktree** — the
Dockerfile copies tracked files by name, so a dirty tree bakes work in
progress into the image.

```bash
git worktree add --detach /tmp/hvtracker-deploy main
cd /tmp/hvtracker-deploy
railway up --detach
```

Deploy from inside the directory: `railway up <path>` given a path outside
the current directory uploads nothing and exits at "Indexing". The Railway
project is named `hvtracker-cron`; the site is the `web` service inside it,
so check `railway status` before assuming. `--ci` exits non-zero on
"Failed to stream build logs" even when the deploy is fine — poll
`railway status` until the deployment id changes rather than trusting that
exit code.

Verify with `curl -A "Mozilla/5.0" https://hvtracker.net/healthz` (Cloudflare
blocks default user agents) and check `agents`, `catalog_agents` and
`render_in_sync` in the response.

---

## Running Locally

```bash
git clone https://github.com/YugantM/hvtracker.git
cd hvtracker
pip install -r requirements.txt

export GITHUB_TOKEN=$(gh auth token)  # or a personal access token
python fetch_and_build.py --render-only

python3 -m http.server 4173
```

Open [http://127.0.0.1:4173](http://127.0.0.1:4173).

### Local Docker Workflow

Use Docker when you want discovery, dry-run additions, renders, and the app
to run in the same local container environment.

```bash
docker compose up --build web
```

This serves the app at [http://127.0.0.1:8080](http://127.0.0.1:8080).

For registry work, run the tooling service against your checked-out repo so
changes to `agents.json` or `candidates.json` persist locally:

```bash
export GITHUB_TOKEN=...  # required for GitHub discovery/API-backed refreshes

docker compose run --rm tooling python discover_agents.py
docker compose run --rm tooling python auto_add_agents.py --dry-run
docker compose run --rm tooling python fetch_and_build.py --render-only
docker compose run --rm tooling python -m pytest tests/test_discovery.py tests/test_data_integrity.py
```

For manual candidate review, edit `docs/import-candidates.json` with either
`owner/name`, full GitHub URLs, or object entries, then query:

```bash
curl -s http://127.0.0.1:8080/api/import-candidates | python -m json.tool
curl -s "http://127.0.0.1:8080/api/import-candidates?status=new&tracked=false" | python -m json.tool
```

Notes:

- `tooling` bind-mounts the repo into `/workspace`, so file changes are written
  back to your local checkout.
- `auto_add_agents.py` writes to `agents.json` only without `--dry-run`.
- `fetch_and_build.py --render-only` stays local and avoids external API calls.

Production runs on Railway with:

- FastAPI for health, API, forms, and dynamic badge routes
- Generated site output stored on a persistent volume
- A 4-hour scheduler: each run fully re-fetches the stalest sixth of the board and refreshes GitHub signals for the rest

---

## Submit Or Correct An Agent

Use the [agent listing issue template](https://github.com/YugantM/hvtracker/issues/new?template=agent-listing.yml).

A listed project should be:

- A public, non-archived GitHub repository with an open-source license
- An AI agent or an agent framework whose primary purpose is agent construction
- Clearly agent-specific in its own codebase, not a general AI SDK, model repo, cookbook, or generic app framework
- Active within the last 12 months
- Not already listed

Include the canonical repository, preferred display name, category suggestion, package names, and any correction details.

For the strict plain-language boundary, see [docs/strict-inclusion-rubric.md](/Users/harsiddhipari/hv_tracker/docs/strict-inclusion-rubric.md).

---

## Specifications

- [Runtime Trust v0.2](https://hvtracker.net/spec/runtime-trust/v0.2/) - the adjustment table the production score is computed from
- [Trust Credential v0.2](https://hvtracker.net/spec/trust-credential/v0.2/)
- [Methodology v2.0](https://hvtracker.net/spec/methodology/v2.0/)
- [Eligibility v1.0](https://hvtracker.net/spec/eligibility/v1.0/)
- [Listing v0.1](https://hvtracker.net/spec/listing/v0.1/)
- [Data Schema v0.1](https://hvtracker.net/spec/data-schema/v0.1/)
- [Provenance v0.1](https://hvtracker.net/spec/provenance/v0.1/)
- [Build Report v0.1](https://hvtracker.net/spec/build-report/v0.1/)

---

## Repository Layout

```text
hvtracker/
├── fetch_and_build.py        # Core build, scoring, and rendering
├── app.py                    # FastAPI edge: serving, /healthz, refresh scheduler
├── auth.py                   # Accounts, watchlist, notifications
├── db.py                     # Postgres layer, falls back to agents.json
├── usage.py                  # Machine-traffic counters behind /live/
├── mcp_server.py             # Hosted MCP server (/mcp)
├── specs.py                  # Specification content
├── template.html             # Main registry template
├── templates/                # Agent, category, blog, compare, and spec templates
├── agents.json               # Curated registry of agents and MCP servers
├── scan_scorecards.py        # Sharded OSSF Scorecard scan
├── discover_agents.py        # Discovery scan
├── scripts/gates.sh          # The PR gates, in one command
├── tests/                    # Suite + validate_html.py
├── seed/history/             # Rank-delta baseline shipped in the image
├── docs/                     # Launch, research, and operating docs
├── data/                     # Generated public data endpoints
├── agents/                   # Generated per-agent pages
├── badge/                    # Generated SVG badges
└── blog/                     # Generated and static articles
```

---

## License

The public data is licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Review [docs/open-core.md](docs/open-core.md) before changing the public/private data boundary for a future company-backed edition.

浏览与调用公开API一般无需安装；本地运行需 Git、Python 与 requirements.txt 中的依赖，以及 GitHub Token；Docker 可选。托管服务的登录、关注列表、私有功能与商业模块费用待核验；第三方MCP客户端连接地址、鉴权方式和传输协议待核验。