TL;DR

• Hermes is a local-first agent framework from Nous Research that creates and refines skills as it works.
• It uses persistent memory to remember user preferences and past conversations between sessions, getting sharper with use.
• Firecrawl is the default web provider, giving the agent search and crawl tools out of the box.

The Hermes agent from Nous Research addresses the biggest limitation of most AI agent frameworks: they're stateless. Every session starts from a blank slate, so you repeat context and the agent never learns anything about you or the tasks you assign it.

Hermes takes the opposite approach. It runs on your own machine, remembers what worked in past sessions, and turns finished tasks into reusable skills it can pull up again later.

Hermes agent is an open-source framework from Nous Research, the lab behind the Hermes and Nomos model families.

The piece that sets it apart from other agent runtimes is the learning loop. When Hermes finishes a complex task, it writes a new "skill" to disk and reuses that skill the next time a similar task comes up.

This learning cycle is built on a flexible architecture. The agent itself can live anywhere you can run a script. It supports six terminal backends including local execution, Docker, SSH, and serverless platforms like Daytona and Modal.

That means you can chat with an agent from a messaging app like Telegram while it works on a remote cloud VM you never log into directly.

The framework is also model-agnostic. You can switch between providers like Nous Portal, OpenRouter, OpenAI, and Anthropic, or connect to a local model through Ollama, by running the hermes model command.Pick the model that fits your budget and task, and swap it out later with one command.

For web access, Hermes uses a pluggable backend system.

Firecrawl is the default, and the framework auto-detects which provider to use from your API keys. Tavily, Exa, and Parallel are also supported. Firecrawl is the only one of the four that covers search, scraping, and multi-page crawling in a single backend.

The core of the agent's persistence is a pair of markdown files.

Hermes stores its own notes and summaries in MEMORY.md, a 2,200-character file, and builds a profile of your preferences in USER.md, a 1,375-character file. Both are injected into the system prompt at the start of every session.

For longer-term recall, the agent uses FTS5 full-text search to query all past conversations and can connect to eight external memory providers like Honcho or Mem0.

When the agent finishes a task that used five or more tool calls, it writes a new "skill" on its own. A skill is a portable knowledge document, compatible with the agentskills.io open standard, that the agent can pull up again the next time a similar task lands.

Skills can also be shared or installed from hubs like ClawHub and the official Skills Hub. Out of the box, Hermes agent ships with 47 built-in tools.

The built-in set covers web search, terminal execution, file editing, and memory management. For bigger jobs, it can delegate work by spawning up to three isolated subagents in parallel.

When the built-in tools aren't enough, MCP servers let it reach GitHub, databases, and other external systems without anyone writing a new native tool.

A built-in cron system lets you schedule automations with no daily run limits, a clear difference from hosted alternatives. The only constraint is your own API budget.

When a scheduled task completes, the agent can deliver the results to over 15 messaging platforms, including Telegram, Discord, Slack, WhatsApp, Signal, and Matrix.

The agent's built-in cron scheduler is the engine behind most automations.

A single command can set up a recurring task, like generating a weekly AI news digest and delivering it to a Telegram chat.

The official documentation provides several automation templates that work with any model, from nightly backlog triage on GitHub to daily arXiv paper summaries saved to Obsidian notes.

These templates show how to combine scheduling with the agent's tools.

An automatic pull request reviewer can be configured to read every new PR and post a review comment directly on GitHub. An uptime monitor checks a list of endpoints every 30 minutes and only sends a notification when a service is down.

Other examples include a nightly bug fix automation that pulls an issue, attempts a fix, and opens a draft PR, or a pricing monitor that watches a competitor's website and notifies you of any change.

An additional example is an weekly AI news digest runs on a schedule and uses the Firecrawl /search endpoint to pull fresh articles on whatever topics you specify before writing and delivering the summary. The agent queries Firecrawl for recent results, gets back clean markdown, and compiles the digest without you touching anything between runs.

Community projects often pair Hermes with other tools. One developer built an eBay deal finder that uses Firecrawl to scrape product listings and prices. Hermes runs on a schedule, pulls the latest data, and decides whether a listing is a good deal against the user's criteria.

The community also maintains DiscoverHermes, a portal for finding new projects, skills, and tools built for the agent.

You can install Hermes agent with a single command piped from its GitHub repository.

The script works on Linux, macOS, Android via Termux, and Windows using WSL2. It's best to include the --skip-setup flag, which keeps the interactive setup wizard from hanging when the script is piped this way.

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash -s -- --skip-setup

The installer clones the project into ~/.hermes/, creates a Python virtual environment, and adds the hermes command to your path.

After it finishes, reload your shell configuration so the new path takes effect.

For Zsh users, that's source ~/.zshrc, and for Bash, it's source ~/.bashrc. With the path updated, run the interactive setup wizard with hermes setup or configure just the model provider with hermes model.

The hermes model command walks you through selecting an inference provider and entering your API key. Hermes works with any model that has at least a 64,000 token context window.

You can connect to a variety of services, each with different model access. All agent settings, including your API keys and configuration choices, are stored locally in the ~/.hermes/ directory.

Once a provider is configured, start a new session by running hermes. The welcome banner shows the agent's version, the active model, and a list of available tools and skills.

From here, you can start a conversation. The agent will respond, and the status bar at the bottom will track token usage and response time for each turn.

Several slash commands help you manage the session. /help lists all available commands, /tools shows the toolset, /model lets you switch models, and /save manually saves the current conversation state. To pick up where you left off in a previous session, use the continue flag: hermes --continue or its shorter version, hermes -c.

Hermes agent's base installation is a starting point. The framework becomes a more practical tool when you connect it to external services and custom logic for memory, skills, and communication.

Adding web access is the most common first step.

Firecrawl gives AI agents fast, reliable web context with strong search, scraping, and interaction tools, and it's the default web backend Hermes ships with. Hermes auto-detects it the moment you set FIRECRAWL_API_KEY in ~/.hermes/.env, and the env var alone is enough (no config edit needed).

If you want to pin the choice explicitly, run hermes tools and pick Firecrawl from the interactive menu. The wizard writes the web: block to config.yaml for you.

When you ask a question that requires current information, the agent decides to use its web_search tool. This triggers a request to Firecrawl's search endpoint, which finds relevant pages, reads their content, and returns clean markdown for the agent to use.

Other backends like Tavily and Exa only handle search, while Firecrawl also does full-page scraping and site-wide crawling.

To point the agent at a self-hosted Firecrawl, set FIRECRAWL_API_URL.

The agent then uses this markdown as context to answer your original question.

Try Firecrawl with Hermes agent. The free tier includes 500 credits. Get your API key and add it to your Hermes setup in under a minute.

Beyond web access, the framework exposes four other extension points that you should be aware of.

MCP servers let Hermes reach tools like GitHub, databases, or local file systems without anyone writing a new native tool. You add them to config.yaml under the mcp_servers key, and the agent gains a new way to interact with external systems.

Custom skills can be written yourself or installed from a hub like skills.sh or ClawHub. The hermes skills search command finds available skills, and hermes skills install adds them to your agent.

Memory providers extend the default memory system. Plugins for Honcho, Mem0, and six other providers are available for more advanced, cross-session user modeling and context retrieval.

Messaging gateways let you interact with your agent from platforms other than the command line. The hermes gateway setup command walks you through connecting to services like Telegram, Discord, Slack, and WhatsApp for receiving notifications or chatting directly with your agent.

Hermes agent is worth picking up if you want an agent that lives on your machine, keeps working between sessions, and doesn't lock you to one model provider. Pair it with Firecrawl for search, scraping, and crawling, and you have an assistant you fully own.

The agent is open source, and the one-line install takes under a minute.

Ready to build your own persistent agent? Get your Firecrawl API key and run the Hermes install script to get started.