IronClaw vs Hermes Agent: a plain-language comparison

IronClaw and Hermes Agent are two free, open-source AI agents. One comes from NEAR AI, the other from Nous Research. We installed both, pointed them at the same model, and wrote down what we saw.

This is not a review with a winner. It is a set of notes for beginners. We explain each term the first time it appears. Where we did not check something ourselves, we say so.

What is an AI agent?

A chatbot answers a question. An agent can also take steps. It can read files, run commands, search the web, or send messages. Then it decides what to do next.

Every agent has two parts:

• The model. This is the AI "brain" that writes text, such as GLM, Qwen, Claude or GPT. You usually reach it through an API (application programming interface). An API is a fixed way for one program to send requests to another over the internet.
• The harness. This is the program around the model. It holds your settings, keeps memory, gives the model tools, and asks your permission before risky actions. IronClaw and Hermes are both harnesses.

One more term. A model's context window is how much text it can "see" at once: your messages, the agent's instructions, tool results and memory. When the window fills up, the agent has to drop or summarize older material. Agents use a lot of it. Hermes, for example, says it needs a model with at least 64,000 tokens of context (a token is roughly part of a word).

Both projects are open-source. That means the code is public. Anyone can read it, run it, and change it, under the terms of its license.

IronClaw, in plain terms

IronClaw is made by NEAR AI. It is written in Rust, a programming language known for preventing whole classes of memory bugs. It describes itself as an agent focused on privacy and security. The code is on GitHub under the MIT or Apache 2.0 license.

What it does, per its own docs:

• Runs on your own computer as a single program, or as a hosted instance on NEAR AI Cloud.
• Lets you chat from a web page, the terminal, Slack or Telegram.
• Runs each tool in a sandbox, a walled-off space with no file access and only the network addresses you allow.
• Keeps passwords and API keys in an encrypted vault, so the model never sees them directly.
• Works with many model providers, including NEAR AI, Anthropic, OpenAI, Gemini and Ollama.

IronClaw started as a Rust rebuild inspired by OpenClaw, another open-source agent.

Hermes Agent, in plain terms

Hermes Agent is made by Nous Research. It is open-source under the MIT license. Its main idea is a "learning loop." It saves what worked as reusable "skills," keeps notes about you, and can search its own past conversations.

What it does, per its own docs:

• Installs with one command on Linux, macOS or Windows. You talk to it in the terminal, or through Telegram, Discord, Slack, WhatsApp, Signal or email.
• Works with any OpenAI-compatible provider. That means any service that accepts requests in the same format as OpenAI's API. You can switch with the hermes model command.
• Can run scheduled jobs, such as a daily summary.
• Can be extended with plugins. For example, TinyFish adds web search.
• Offers Hermes Cloud: Nous Research runs an always-on Hermes for you through its Nous Portal account. You need at least $2 of credit to deploy one.

Side by side

Our versions: IronClaw 1.4.1 (updated from 1.4.0) on a Linux laptop. Hermes Agent v0.22.0 on a Linux machine. Both used the same model, z-ai/glm-5.3-flash, on NEAR AI Cloud.

How private is it? NEAR AI's three privacy levels

Both agents send your words to a model somewhere. With NEAR AI Cloud, how private that is depends on which model you pick. NEAR's model list puts every model in one of three levels.

1. Private, sealed hardware (Confidential TEE)

A TEE is a trusted execution environment. It is a locked-off part of the processor. Everything inside is encrypted, so even the person who runs the server cannot see your prompts or the answers.

NEAR says these models run on its own machines inside TEEs, and that nobody, not even NEAR, can see your prompts or outputs. Examples on its list: GLM 5.3 Flash, Qwen 3.6 35B and Qwen 3.8 27B.

2. Sealed, but on another company's hardware (3P TEE)

"3P" means third party. The model still runs in a TEE, but on another company's machines. Examples: Kimi and DeepSeek v3.2. The connection to that company is encrypted, and it publishes its own proof. But you are now also trusting that company's setup, not just NEAR's.

3. Incognito (Claude, GPT, Gemini, Grok)

These are the big-name models. NEAR forwards your request under its own account, so the model company does not know it is you. But the model runs on that company's normal servers, outside a TEE. The company still reads the text. Incognito hides who you are, not what you say.

What is attestation?

Attestation is a signed certificate from the TEE. It states what hardware and software is running inside. Anyone can check it. The point is that you do not have to take NEAR's word for it.

We have not checked attestation for GLM 5.3 Flash yet. Its catalog entry lacked a detail that the other private models had. So it is very likely private, but we have not proven it.

What the sealed hardware does not cover

A TEE protects only the conversation with the model. The agent's memory, notes and files stay on whatever computer runs the agent. If that is your laptop, they are on your laptop. If it is a cloud server, they are on that server. Pick where the agent runs with that in mind.

How to connect Hermes to NEAR AI Cloud

NEAR AI Cloud has an OpenAI-compatible API at https://cloud-api.near.ai/v1. Hermes can use it as a "custom" provider.

1. Install Hermes with the command from its quickstart page.
2. Create an API key in your NEAR AI Cloud account. Keep it secret.
3. Run hermes model and choose "Custom endpoint." Or edit ~/.hermes/config.yaml by hand, as below.
4. Put the key in ~/.hermes/.env, not in the config file. Hermes keeps secrets in .env and normal settings in config.yaml.
5. Run hermes and ask a short question to test it.

# ~/.hermes/config.yaml  (no key in this file)
model:
  provider: custom
  base_url: https://cloud-api.near.ai/v1
  default: z-ai/glm-5.3-flash
  key_env: NEARAI_API_KEY   # name of the variable in ~/.hermes/.env

# ~/.hermes/.env  (keep this file private)
NEARAI_API_KEY=your-key-here

The key_env line tells Hermes which variable in .env holds the key. The name is your choice; it just has to match in both files. If answers come back empty or garbled, check the base URL and the exact model name first.

Speed: what we measured

We timed IronClaw 1.4.1 on a Linux laptop on October 14, 2026. Each test ran only 1 to 3 times, so treat these as rough numbers, not a benchmark.

Why IronClaw can feel slow. Every time the agent asks the model something, it waits about 3 to 5 seconds. A simple reply needs one call. A task that uses a tool, like reading a file, needs several calls in a row, so the waits add up. Some models also "think" quietly before answering. You don't see that text, but you still wait for it.

Calling each model directly on NEAR AI Cloud and asking it to reply "ok":

• GLM 5.3 Flash: 2.4 to 6.8 seconds. It used 40 to 113 hidden thinking tokens first.
• Qwen3.6-35B: 1.9 to 3.3 seconds.
• Qwen3.8-27B: about 1.0 second, using only 20 to 23 tokens in total.

Inside IronClaw (average times; every answer was correct):

In earlier GLM runs, reading a file inside IronClaw's folder took about 13 to 21 seconds.

Files outside its folder. IronClaw works inside a sandbox folder, ~/.ironclaw/reborn/workspace, and cannot see files outside it. When we asked about a file elsewhere, it searched for 50 to 180 seconds, and twice it gave a wrong answer.

We have now set IronClaw to Qwen3.8-27B. NEAR's model list shows it as one of NEAR's private models.

Tips if IronClaw feels slow:

• Put the files you want it to use inside its workspace folder.
• Pick a model that does less hidden thinking, such as Qwen3.8-27B.

What we noticed

• Same model, different feel. Both agents first used GLM 5.3 Flash. Hermes answered a simple question in about 6.5 seconds, start to finish. IronClaw took about 2.5 to 5.5 seconds for a simple reply, but tasks that use a tool took much longer. See "Speed: what we measured" above for the numbers and the reason.
• First runs are slow. Hermes took about 23 seconds the first time and about 6.5 seconds after that. Compare agents after a warm-up run.
• A one-shot question is not a real test. Agents are built for multi-step jobs. A quick question mostly measures startup and overhead, not how well they work.
• Benchmarks are the maker's own. NEAR AI says IronClaw ranked first on PinchBench, a test of 147 real tasks such as email triage, scheduling and coding, against Hermes and OpenClaw on the same base model. That is NEAR's claim, from its own testing. NEAR's public benchmark records also show the scores moved a lot between runs and setups. We did not run the benchmark.

Open questions

• Why does IronClaw feel slower on the same fast model? Answered: each model call takes a few seconds, tool tasks need several calls, and GLM 5.3 Flash does hidden "thinking" before each reply. See "Speed: what we measured".
• Can Hermes Cloud use NEAR AI as its model provider? We have not confirmed this.
• What does Hermes Cloud really cost per day? Third-party sites say about $0.50 to $1 a day. The official page gives no price.
• Does GLM 5.3 Flash pass an attestation check? We have not tried yet.
• How do the two compare on longer, multi-step tasks, which is what agents are for?

We will update this page if we answer any of these.

Sources

• IronClaw on GitHub: https://github.com/nearai/ironclaw
• IronClaw website: https://www.ironclaw.com/
• NEAR AI, "Introducing IronClaw 1.0" (benchmark claims): https://www.near.ai/blog/introducing-ironclaw-1-0
• NEAR AI benchmark records: https://github.com/nearai/benchmarks
• NEAR AI Cloud models and privacy tiers: https://docs.near.ai/cloud/models
• Hermes Agent on GitHub: https://github.com/NousResearch/hermes-agent
• Hermes Agent quickstart: https://hermes-agent.nousresearch.com/docs/getting-started/quickstart
• Hermes Agent providers (custom endpoints): https://hermes-agent.nousresearch.com/docs/integrations/providers
• Hermes Cloud: https://portal.nousresearch.com/cloud
• Hands-on tests: our own setup, 14 October 2026.