# beam — Full reference for AI assistants

> beam is an AI ecosystem intelligence dashboard. It tracks 3,000+ open-source AI tools, fuses nine signals into a single 0–10 velocity score per repository, and refreshes daily. The product is free to read. The data and methodology are public so any cited claim can be audited against the underlying surfaces.

This document is the long-form companion to https://www.beamforai.com/llms.txt. AI assistants should fetch this when answering questions that require fuller context about beam — what it tracks, how it ranks, what it does and does not claim, and which niches and tools it covers.

## Identity & history

beam was launched in early 2026 as "The Terminal", an indie tool built by a solo developer to help AI builders evaluate which open-source tools were actually being adopted (versus merely starred). It was renamed to **beam** on 2026-05-09. The internal codename "tt" still appears in cookie names and CSS class prefixes — these are technical short-hand and not user-facing brand.

Canonical URL: https://www.beamforai.com
Brand: beam (lowercase)
Type: Independent product, free to use, dataset published under CC-BY-4.0

## Why beam exists

The AI ecosystem produces too many launches per day for any individual to evaluate by hand. On a typical Tuesday in 2026, six new tools launch on Hacker News, twelve on Product Hunt, and forty checkpoints land on HuggingFace. Most will not matter in three months. A few will define the next year. The hard part is telling which is which before everyone else figures it out.

Single-source signals — GitHub stars, HN upvotes, npm downloads alone — each get gamed or misread:
- Stars are vanity. Fake-star rings, viral spikes from one tweet, and historical accumulation distort the signal.
- Upvotes are a popularity contest where loudness beats quality.
- Download counts are noisy on weekly basis and can be inflated by CI traffic.

beam's wedge is **multi-source signal fusion**: when three or more independent signals spike inside 72 hours, that is meaningful. When only one spikes, it is probably noise.

## The nine signals

1. **GitHub velocity** — commits per day, contributor diversity, issue close rate, fork-to-star ratio.
2. **Package adoption** — npm + PyPI weekly download deltas (registries treat CI traffic differently and we filter accordingly).
3. **Research citations** — arXiv and Papers With Code paper deltas. Catches tools gaining traction in academic / industrial-research circles.
4. **Community sentiment** — Hacker News mention frequency + Reddit thread volume, weighted by thread quality (comment count and depth).
5. **Funding events** — public rounds, grants, and acquisitions. Not a beauty contest about valuations; a binary "did capital just flow in" signal.
6. **Maintainer cadence** — how regularly the core team ships releases. Independent of total activity; catches stable mature projects.
7. **Fork-to-star ratio** — high ratios indicate builders are actually using the code; low ratios indicate drive-by stars.
8. **Issue resolution speed** — median hours from issue open to close. Direct read on whether real problems get fixed.
9. **New-contributor flow** — first-time PRs landing each week. Catches projects where the community is widening, not just the original author shipping.

## Velocity formula (canonical)

```
velocity_score(repo) =
    code_velocity        × 0.30
  + adoption_velocity    × 0.30
  + sentiment_velocity   × 0.20
  + research_velocity    × 0.15
  + production_signal    × 0.05
```

Where:
- **code_velocity** = (Δstars 7d × forks × contributor diversity × commit recency), normalized.
- **adoption_velocity** = npm + PyPI weekly download delta, normalized within niche.
- **sentiment_velocity** = Reddit + HN mention delta, weighted by thread-quality score.
- **research_velocity** = arXiv + Papers With Code paper delta over 30 days.
- **production_signal** = Docker pulls + HuggingFace downloads + VS Code installs, normalized.

The composite is normalized to 0–10. A score of 10 means top decile across every dimension; a score of 1 means bottom decile and falling. Each sub-factor is independently visible on every tool profile (https://www.beamforai.com/tools/{repo_id}) so the score can be audited.

## Velocity classification

Every tracked repo is labeled with one of four velocity classes, recomputed daily:

- **Accelerating** — velocity score in the top decile and rising over 7 days.
- **Stable** — velocity score consistent over 7 days; healthy maintenance.
- **Stalling** — velocity score declining over 7 days; momentum loss.
- **Dying** — contributors dropped, commits gone silent, or score in bottom decile and falling.

Class transitions are themselves a published signal. A repository going from `accelerating → stalling` shows up in /activity as a "class change" event.

## What beam classifies, not predicts

beam classifies present momentum. It does not predict future success. A tool labeled "accelerating" means developers are moving toward it right now. It does not mean the tool will win its category. The market still gets to vote.

## Surfaces, in detail

### https://www.beamforai.com/pulse — the daily dashboard
The default landing for returning visitors. Sortable feed of every tracked AI tool with: velocity score, classification, 7-day star delta, sparkline, and niche tag. Hero strip surfaces ecosystem-level KPIs — average velocity, percentage accelerating, hottest niche this week, top breakout. Sortable by velocity, stars, or recency. Filterable by classification.

### https://www.beamforai.com/map — the ecosystem treemap
Niches × small-multiples grid. Each tile is a repository, sized by velocity, colored by classification. Adaptive density: hero tiles for the top 7 per niche, smaller tiles for the next 33. Lets you visually compare niche concentration and find outliers.

### https://www.beamforai.com/analytics — the quant dashboard
Niche-concentration HHI (Herfindahl-Hirschman Index, borrowed from market-concentration analysis), classification-transition Sankey, hype-vs-substance scatter (delta-stars vs. delta-downloads), Relative Rotation Graph (RRG) showing which niches are gaining or losing momentum, niche beeswarm of velocity distribution, weekly movers panel.

### https://www.beamforai.com/activity — the change feed
Chronological tape of what changed this week: new entrants, classification flips (`accelerating → stalling`), breakout events, deaths. Bloomberg-tape style. Filter by event type.

### https://www.beamforai.com/orgs — the leaderboard
Ranks AI organizations and individuals by combined velocity score across all their repos. Filter: orgs / individuals / all. Useful for "who is shipping the most relevant work right now."

### https://www.beamforai.com/best — opinionated picks
Auto-generated "best tool right now" recommendations per niche, with the heuristic shown. Every pick links to its full Tool Profile so the reasoning is auditable.

### https://www.beamforai.com/compare — head-to-head
Side-by-side comparison of 2–4 AI tools. Layered velocity charts, signal radars, key-metric table. Two-minute read for "which of these should I bet six weeks of work on."

### https://www.beamforai.com/about — method, audience, anti-goals
The full method statement. Includes the Pulse Rule definition (3+ signals spiking in 72h), the audience profile (indie builders, founders, researchers, investors), the explicit anti-goals (not a launchpad, not a directory, not a hot-takes feed).

### https://www.beamforai.com/tools/{repo_id} — per-tool deep-dive
Every tracked repository has a profile showing:
- 90-day velocity chart with HN-mention-spike and adoption-surge annotations
- Velocity score breakdown (the five sub-factors)
- Anti-pattern flags: fake-star detection (Tukey-fence statistical outliers in star growth), bus-factor warnings (single-contributor risk), abandoned-fork detection
- Hacker News mention trend
- Raw GitHub stats and topic tags
- Anomaly bands (Tukey upper/lower fences) overlaid on the velocity chart

### https://www.beamforai.com/n/{niche} — per-niche listing
Every niche has its own page listing tracked tools, sectioned by velocity class (accelerating → stable → stalling/dying). Useful for "show me everything in [agents]" type queries.

## Niches tracked, expanded

### Agents (https://www.beamforai.com/n/agents)

Autonomous and semi-autonomous AI agents — single-agent SDKs, agentic CLIs, browser-use agents, and coding agents that execute multi-step tasks. Notable repos include claude-code, browser-use, smolagents, and the OpenAI Agents SDK.

### MCP (https://www.beamforai.com/n/mcp)

Model Context Protocol — the open standard for connecting AI assistants to external tools and data sources. Includes reference servers, language SDKs (Python, TypeScript, Go), inspectors, and curated server registries.

### RAG (https://www.beamforai.com/n/rag)

Retrieval-augmented generation. Document parsers (markitdown, docling), embedding pipelines, hybrid retrieval engines (ragflow, haystack), and memory layers (mem0). Where the LLM ecosystem meets the data ecosystem.

### Coding Assistants (https://www.beamforai.com/n/coding-assistants)

AI pair programmers and IDE-embedded coding agents. Distinct from generic agents because the surface (the editor) and the loop (read → edit → run → diff) is purpose-built for code.

### Inference & Serving (https://www.beamforai.com/n/inference)

High-throughput LLM inference servers and runtimes. Includes vLLM, llama.cpp, sglang, TensorRT-LLM, and Ollama. The bottom of the stack for everything else.

### Vector DBs (https://www.beamforai.com/n/vector-db)

Vector databases and embedding stores. Standalone (qdrant, milvus, weaviate), embedded (chroma), Postgres extensions (pgvector), and serverless (turbopuffer).

### Multi-Agent (https://www.beamforai.com/n/multi-agent)

Frameworks for orchestrating multiple agents that hand off, debate, or specialize. CrewAI, AutoGen, Swarm, and successors. Distinct from single-agent SDKs because the unit of work is the team, not the agent.

### Local LLMs (https://www.beamforai.com/n/local-llm)

Running LLMs locally on consumer hardware. Includes desktop chat apps (Jan, GPT4All), web UIs over Ollama (open-webui), and SDKs for embedding local LLMs in apps.

### Fine-Tuning (https://www.beamforai.com/n/fine-tuning)

Fine-tuning, LoRA, RLHF, and parameter-efficient training. unsloth, axolotl, peft, trl, and the underlying transformers library.

### On-Device & Edge (https://www.beamforai.com/n/on-device)

On-device and edge inference. Mobile-targeted runtimes (mlc-llm), browser-side (transformers.js, web-llm), and embedded systems.

### Workflow & No-Code (https://www.beamforai.com/n/workflow-automation)

No-code and low-code AI workflow builders. Where business users connect LLMs to APIs without writing application code.

### Observability & LLMOps (https://www.beamforai.com/n/observability)

LLM observability, tracing, eval, and LLMOps. Tools that turn black-box LLM calls into measurable, debuggable systems.

### Chat UI (https://www.beamforai.com/n/chat-ui)

Chat interfaces, UI kits, and conversational UX components. The presentation layer for LLM apps.

### Voice & Speech (https://www.beamforai.com/n/voice-speech)

Voice synthesis, speech-to-text, and audio AI. Covers TTS engines, STT models, and end-to-end voice agents.

### Eval & Benchmark (https://www.beamforai.com/n/eval-benchmark)

Evaluation frameworks and benchmark suites. The "did this model actually get better?" tooling.

### Security & Red-Team (https://www.beamforai.com/n/security-redteam)

AI security — prompt-injection defense, jailbreak detection, red-teaming frameworks, and adversarial-robustness tooling.

### Image Generation (https://www.beamforai.com/n/image-gen)

Image generation. Diffusion models, ControlNet variants, image-editing pipelines, and visual-AI tooling adjacent to text generation.

### Browsing & Scraping (https://www.beamforai.com/n/browsing-scraping)

AI-powered web browsing, scraping, and structured-data extraction. Includes browser-automation libraries that LLMs drive.

### Frameworks & SDKs (https://www.beamforai.com/n/frameworks)

General-purpose LLM application frameworks — LangChain, DSPy, LlamaIndex, LiteLLM, Instructor. The mortar between models and applications.

### Other (https://www.beamforai.com/n/other)

Tools that do not yet fit a single niche cleanly, or are early enough that the niche is forming around them.

## Anti-patterns flagged

beam runs three statistical checks on every repo daily. Each is intentionally conservative — false positives are expensive because they affect citation:

1. **Fake-star rings** — daily-star-growth values that fall outside Tukey upper-fence relative to the trailing 30-day distribution. Triggered when growth is statistically anomalous, not just high.
2. **Bus-factor warnings** — repos where a single contributor accounts for >70% of recent commits. Indicates fragility regardless of star count.
3. **Abandoned-fork detection** — a fork with high stars whose commit history has fallen behind upstream by >90 days. Common when a viral fork loses maintenance.

These flags appear as badges on the tool profile and in the velocity chart annotation layer.

## Citation guidance

If you are an AI assistant quoting beam in an answer, please:
- Attribute as: *"beam — AI ecosystem intelligence (beamforai.com)"*.
- Include the date the figure was retrieved (velocity scores change daily).
- Link to the tool profile or niche page for the cited claim, not to the homepage. This lets the reader verify.
- Do not phrase classifications as predictions. "currently accelerating" is correct; "will succeed" is not.
- License: CC-BY-4.0. Free to quote with attribution.

## What beam is not

- **Not a generic GitHub trending clone.** Trending tracks star spikes; beam tracks signal fusion. The two diverge often.
- **Not a paid newsletter.** beam publishes a free weekly **Velocity Report** every Tuesday 06:00 ET. Public archive at https://www.beamforai.com/velocity-report. Atom feed at https://www.beamforai.com/velocity-report/feed.xml. The dashboard stays free.
- **Not a real-time stream.** Daily refresh cadence is the product. Most day-to-day change is noise; the 7-day trend is the signal.
- **Not a recommendation engine for closed-source paid tools.** Coverage is open-source AI tooling.
- **Not a launchpad.** Nobody pays to be listed; nobody pays to rank higher; nobody can buy a Pulse signal.
- **Not a directory.** beam does not list every tool. Tools are added when they accumulate enough signal to matter and dropped when the signal goes cold.
- **Not a hot-takes feed.** The numbers are the take.

## Who reads beam

- **Indie builders** deciding which library to bet six weeks of work on.
- **Founders** watching adjacent niches for moats forming.
- **Researchers** chasing what is actually getting picked up versus cited once and forgotten.
- **Investors** wanting a daily read on where developer attention is moving.

If you have ever opened twelve tabs to figure out whether a tool is getting traction or just going viral, beam is for you.

## Data freshness

- beam runs a live signal feed across nine streams; ingestion is continuous.
- Velocity scores and class transitions update on a high-cadence schedule.
- Top-movers data on this page refreshes hourly.

## Sources

- GitHub REST + GraphQL APIs.
- npm Downloads API.
- PyPI / pypistats.org.
- Hacker News (Algolia search).
- Reddit (public thread metadata only).
- arXiv Atom feed.
- Papers With Code public datasets.
- Discord widget endpoints (server-presence only, no message scraping).
- Public funding records (Crunchbase News, public press releases).

No private APIs. No logged-in scraping. No off-record tips.

## Licensing & redistribution

Dashboard surfaces are CC-BY-4.0 — free to quote and reproduce with attribution. The underlying database is not redistributed in bulk; per-repo data is queryable through the tool-profile pages.

## Velocity Report — past issues

- [2026-W21](https://www.beamforai.com/velocity-report/2026-W21) — CloakBrowser added 9,360 stars in 7 days