Hermes Atlas: a curated map of the Hermes Agent ecosystem

What Hermes Atlas actually solves

Hermes Agent, the self-improving agent from Nous Research, launched in February 2026, and the README says it "immediately spawned a fast-growing community of skills, plugins, integrations, deployment templates, and forks." Discovery is the problem that follows: those projects are spread across GitHub, and a plain search returns everything, including abandoned experiments. Hermes Atlas is a directory that answers a narrower question: which Hermes Agent projects are worth looking at. Its contribution is editorial plus mechanical. Every entry passes a stated filtering rule set (built for or integrated with Hermes Agent, created after July 22, 2025, not a personal pet project or assignment, and a basic security review), and the map is then kept current by pulling star counts from the GitHub API. The audience is someone who has decided to use Hermes Agent and now needs to pick skills, plugins or deployment templates without reading fifty repositories. It is not for someone who has not chosen an agent yet; the whole site assumes Hermes Agent is the substrate.

How the map, the star API and the chatbot fit together

The architecture is deliberately small. index.html is a single-page app in vanilla HTML, CSS and JavaScript with no framework and no build step, served as a static file from Vercel. Around it sit three serverless functions: /api/stars fetches GitHub star counts and caches them in Redis Cloud with a one-hour TTL, /api/stars-history serves 30-day history for the sparklines, and /api/chat runs the retrieval-augmented chatbot. The data flow for stars is index.html to /api/stars to Redis to the GitHub GraphQL API, and the README notes that the GitHub token raises the rate limit from 60 to 5000 requests per hour. The chatbot is the more involved path. Research files are split into chunks and embedded with OpenAI text-embedding-3-small at build time, cached as a static JSON file plus a float32 embeddings.bin of roughly 14MB, and retrieved at query time with hybrid BM25 and cosine similarity, MMR re-ranking, and conversation-aware query rewriting before the prompt reaches OpenRouter. The README describes a fallback chain across models, and the environment table allows overriding the primary model with OPENROUTER_MODEL and the chain with OPENROUTER_FALLBACK_MODELS. One inconsistency is worth flagging: the stack section names Gemma 4 models, while the environment table lists the default as deepseek/deepseek-v4-flash with Gemini fallbacks. A reader deploying this should treat the environment table as the operative default.

Running Hermes Atlas locally and rebuilding the knowledge base

The README is explicit that local use is partial. Cloning and installing dependencies is only needed for the API endpoints and the chunk builder, and the static site can be previewed by opening index.html directly. The API endpoints, however, only work when deployed to Vercel, so a local checkout gives you the map and the build scripts but not a working /api/stars or /api/chat.

git clone https://github.com/ksimback/hermes-ecosystem.git
cd hermes-ecosystem
npm install

After editing anything under research/, the chatbot's knowledge base has to be rebuilt, which calls the embedding API and therefore needs a key.

OPENROUTER_API_KEY=sk-or-... node scripts/build-chunks.js

The README also provides a local quality test for the retrieval pipeline, which it says passes 27 of 27 cases.

OPENROUTER_API_KEY=sk-or-... node scripts/test-rag.js

For a real deployment, the environment table lists the variables to set in Vercel: GITHUB_TOKEN for the higher rate limit, OPENROUTER_API_KEY and REDIS_URL as required, and the two optional model overrides. Expect the first page load to be fast because the map is static, and the chatbot to be the only part that depends on live keys.

Where the directory model breaks down

The strongest limitation is structural: a curated directory is only as current as its last edit. Star counts refresh hourly, but inclusion is manual, and the README routes new entries through a GitHub issue. There is no documented automated ingestion, so a project that appears next week is invisible until someone opens that issue. The filtering rules also create blind spots by design. Requiring creation after July 22, 2025 excludes older Hermes-adjacent tooling, and excluding personal pet projects and assignments removes early experiments that sometimes become the useful ones. The security review is described as basic, which is a fair description of what a directory can do; it is not an audit, and the README does not claim otherwise. Two smaller gaps matter for operators. The README does not document rollback for a bad cached star count, so a stale Redis entry persists until the TTL expires. And the licence field on the repository is unknown despite the README stating MIT for site code and CC BY 4.0 for research content, which is the kind of mismatch that should be resolved before anyone reuses the research files. If you need an agent runtime, an SDK, or a plugin loader, this is the wrong repository entirely: it contains none of those.

Hermes Atlas compared with an awesome-list or a raw GitHub search

The obvious alternative is a community awesome-list, and the difference is maintenance mechanics rather than philosophy. An awesome-list is a Markdown file updated by pull request; its entries are static text, and any freshness signal (stars, last commit) is stale the moment it merges. Hermes Atlas keeps the same editorial gate but adds a serverless layer that re-reads star counts on a one-hour TTL and stores 30-day history for sparklines, so the map shows movement rather than a snapshot. The second alternative is searching GitHub directly, which has no editorial gate at all and no category structure; Hermes Atlas organizes entries across 12 categories. The trade-off runs the other way too: an awesome-list is trivially forkable and readable in a terminal, while Hermes Atlas depends on Vercel, Redis Cloud and an LLM key for its two dynamic features. If you want a file you can grep, ECOSYSTEM.md in the repository is described as a Markdown version of the map, which is the closest thing to that workflow here.

Maintenance cost, licensing and what to check before adopting

For a reader who only browses hermesatlas.com, the maintenance cost is zero. For anyone self-hosting or forking, the cost sits in three external services. Redis Cloud holds the star cache and the daily history snapshots, so the cache is a hard dependency for /api/stars and /api/stars-history. OpenRouter serves the chatbot, and the README's fallback chain exists precisely because a single provider can fail. The GitHub token is optional but changes the ceiling from 60 to 5000 requests per hour, which matters if the map grows. The embeddings are computed once at build time and committed as static files, so re-embedding is a manual step after editing research/ rather than a runtime cost. On licensing, the README states MIT for site code and CC BY 4.0 for research content, and notes that repository descriptions and metadata come from the upstream projects' own documentation; the repository's licence field is unknown, so anyone reusing the research content should confirm the terms directly rather than assume. The README also states the project is not officially affiliated with Nous Research.

Editorial conclusion

Adopt Hermes Atlas if you are new to Hermes Agent and want a filtered starting point rather than a raw GitHub search, and treat the repository as a static site plus three serverless functions rather than as infrastructure. Do not adopt it if you need an agent runtime, an SDK, or a plugin loader: the repository contains no agent code, and the README describes it as a directory, not a framework. Before relying on it, verify what the README leaves open: the licence field is unknown even though the README states MIT for site code and CC BY 4.0 for research content, the default model in the environment table (deepseek/deepseek-v4-flash) differs from the stack section (Gemma 4), and the README does not document rollback when a cached star count is wrong. Check data/repos.json against the live site before you cite any entry.

Frequently asked questions

Official sources

1. Issues
2. ksimback/hermes-ecosystem on GitHub
3. Project website
4. README