You can run Hermes Agent end-to-end for $0 per token today: self-host the MIT-licensed agent on any Linux box (or your old laptop), point it at the Nous Portal Free tier (which exposes $0 models like Tencent Hy3 and StepFun Step 3.7 Flash), and you have a continuously-running, memory-keeping, tool-using agent that costs you nothing to keep alive. The catch most people hit has nothing to do with the bill — it is that they set the agent up like a chatbot, with one vague prompt and no memory, and it stalls 20 minutes into any real job. The fix is a three-pillar setup — Job, Tools, Memory — plus free-model selection that matches the workload.
Last verified: July 23, 2026
- The agent itself: Hermes Agent by Nous Research, MIT license, $0 to self-host — runs on Linux, macOS, or Windows/WSL2.
- Free model lane: Nous Portal Free tier ($0/mo) includes $0 models such as Tencent Hy3 (Hunyuan 3), StepFun Step 3.7 Flash, and Poolside Laguna S 2.1.
- Second free lane: OpenRouter's
tencent/hy3:freeran to July 21, 2026; standard paid endpoint now$0.14/M input · $0.58/M output. Other $0 OpenRouter endpoints remain.- Third free lane: Local models via Ollama or LM Studio — fully offline, no rate limits.
- Volatile facts: Free-model lists, pricing, and rate limits change often; re-verify before committing.
What is Hermes Agent (and why "free" is even possible)?
Hermes Agent is an open-source, self-improving AI agent from Nous Research. Unlike ChatGPT or Claude in a browser, Hermes runs as a server-resident process — it stays alive, keeps memory between sessions, can be triggered by Telegram/Discord/Slack/webhooks, schedules its own jobs via cron, and learns reusable skills from experience so it never has to solve the same problem twice. The source repo is at github.com/NousResearch/hermes-agent. v0.19 (codename "Quicksilver") shipped mid-2026; see our Hermes Agent 0.19 Quicksilver upgrade guide for the changelog.
The reason a free stack is possible at all is simple: Hermes is provider-agnostic. The agent's intelligence is a pluggable LLM endpoint — you can pull from Nous Portal, OpenRouter (200+ models), direct provider keys (OpenAI, Anthropic, Google, z.ai/GLM, local Ollama), or any OpenAI-compatible endpoint. None of those choices lock you in. The MIT license means the framework costs nothing; the bill you pay is only the model tokens. So "run Hermes for free" reduces to "pick a $0 model endpoint and install the agent."
This is the same architecture shift we covered in how to build an AI agent team with Hermes Agent — the agent loop (planning → tool call → result → next step) is decoupled from the model that powers it.
The three free lanes — pick one (or stack all three)
You have three legitimate, primary-source-verified paths to put a $0 model behind Hermes. Each has different cost, rate limits, and offline characteristics.
| Lane | Cost | Setup | Best for | Limits |
|---|---|---|---|---|
| Nous Portal Free tier | $0/mo, $0 credits | hermes setup --portal and OAuth |
Real agent work with $0 models (Hy3, StepFun, Laguna) | Standard rate limits; free models only; no Tool Gateway |
| OpenRouter free endpoints | $0/token | Sign up at openrouter.ai, paste key, pick a :free model |
Trying new models, low-volume jobs, fallback routing | Per-model RPM/RPD caps; ~200 req/day typical |
| Local models (Ollama / LM Studio) | $0 + your hardware | Install Ollama, hermes model, pick a local model |
Privacy, offline, unlimited runs, sensitive data | Bounded by your GPU/CPU; quality ceiling lower |
Nun? Nous. A 20-second primer on the Portal
The Nous Portal Free tier is the cleanest of the three. Per the official pricing page: Free costs $0/month, includes free models only (currently Tencent Hy3, StepFun Step 3.7 Flash, Poolside Laguna S/XS 2.1), and standard rate limits. It does not include the Tool Gateway's hosted web search, browser automation, FAL image gen, or Firecrawl scraping — those bill to the credit balance and only paid tiers ($20 Plus up to $200 Ultra) include hosted tool usage. That is fine for most agent jobs: Hermes has 40+ built-in tools (web search, browser automation, image gen, terminal, file I/O, calendar, MCP client) that work without the Portal gateway.
Why Tencent Hy3 is the right free model for agent work
Hy3 (also branded Hunyuan 3.0) was released by Tencent on July 6, 2026 — a 295-billion-parameter Mixture-of-Experts model with only 21B parameters active per token, a 256K-token context window, Apache 2.0 licensing, and weights on HuggingFace as tencent/Hy3-preview. Per the Tencent model card and third-party reviews (Caixin Global, July 6 2026; aitooltier.com), the model's post-training was tuned specifically for tool calling, long-context reasoning, and agent tasks — exactly the workload that breaks weaker free models.
Vendor-reported benchmarks (not independently reproduced — label these as Vendor claim):
| Benchmark | Hy3 score | Source |
|---|---|---|
| GPQA Diamond | 90.4 | Tencent model card |
| SWE-bench Verified | 78.0 | Tencent model card |
| SWE-bench Pro | 57.9 | Tencent model card |
| IMOAnswerBench | 90.0 | Tencent model card |
Why these numbers matter for agents: GPQA Diamond tests graduate-level reasoning under pressure; SWE-bench Verified tests real coding task completion against held-out test cases; both are proxies for "can this model hold a long, multi-step job without falling apart." That maps directly to the 20-minute job test (below) — a long task is where weak free models fail and where Hy3 is built to survive.
For context on where Hy3 sits in the open-weights landscape, see our open-weight vs closed-weight AI 2026 breakdown.
How do you install and run Hermes Agent for free?
Run these four steps. They take about 15 minutes the first time.
Step 1 — Install Hermes Agent (one command)
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
This pulls uv, Python 3.11, and the agent itself. Works on macOS, Linux, and WSL2 on Windows. After install, hermes --version should print the build number.
Step 2 — Plug in a free model lane (pick one)
Option A — Nous Portal Free (no card, $0/model):
hermes setup --portal
# Follow the OAuth sign-in. Free tier is selected automatically.
hermes model # confirm a free model is set, e.g. tencent/hy3
Option B — OpenRouter free endpoints:
# 1. Sign up at openrouter.ai, copy your API key.
# 2. Configure Hermes to use OpenRouter:
hermes setup # pick OpenRouter, paste key
# 3. Pick a $0 model:
hermes model tencent/hy3:free # window may have closed 2026-07-21 — pick any *:free model
Option C — Local via Ollama (offline, no rate limits):
# Install Ollama (one command from ollama.com), then:
ollama pull qwen3:32b # or any open-weights model your hardware fits
hermes model # pick the local Ollama endpoint
If you want a deeper walkthrough of local model selection (including how to fit a 70B-class model into 24GB VRAM), we walk through it in Gemma 4 free local coding assistant and 10 free open-source AI tools you can run yourself.
Step 3 — Run the doctor (sanity check)
hermes doctor
hermes doctor checks that the agent process, model provider, memory DB, and gateway are all wired correctly. If anything is broken, it tells you the exact fix. Run this before your first real job — most "the agent isn't doing anything" reports are missing config the doctor would have caught.
Step 4 — Send a real prompt
hermes
> Read this week's notes from ~/notes/*.md, group them by topic, count
mentions per topic, list the top 5, and write a one-sentence takeaway
for each. Save to ~/notes/weekly_digest.md.
If the model + setup is sound, you will watch the agent read files, group them, and produce a digest. If it stalls after step one — read on.
Why free models break agents — and how the Job/Tools/Memory model fixes it
Most people treat an AI agent like a smarter chatbot. They open a window, type a vague request, get an answer, close it, and lose everything. That is a search bar with better manners. Hermes runs differently because every job is built on three pillars. Strip the agent down and these three are all there is.
Pillar 1 — The Job (the instruction set, written down)
A job is not "help me with marketing." A job is a written, reusable instruction block that says what to open, what to look for, what to output, and where to save it — as if you were onboarding someone on their first day.
Vague (broken):
Summarize our community questions.
Good (works):
Read every message in
~/community/questions.md. Group questions by topic. Count how many mentions each topic has. List the top five topics. For each, write a single sentence describing what people are confused about. Save the result to~/community/weekly_questions.md.
Same task, completely different result. The fix is mechanical: name the file, name the operation, name the output, name the destination.
Pillar 2 — The Tools (and only the tools the job needs)
An agent that has 15 tools and a vague job will pick tools at random and burn 20 minutes doing nothing useful. Start narrow — three tools, get the job working, then add. Hermes lets you scope tools per profile or per job; use that scope. Most community-question-sorting jobs need only: read-file, write-file, web-search. That is three tools.
Pillar 2 is also where the model quality shows up, because the model is what hits the tools correctly. A weak model picks the wrong tool, sends the wrong arguments, gets a result back, and forgets what it was doing. A tuned-on-tool-calling model like Hy3 (per its Tencent model card) or a paid frontier model like Claude Opus is what makes the loop hold together.
Pillar 3 — Memory (cross-run state)
Without memory, an agent that sorted last week's questions starts from zero this week. With Hermes memory, it remembers last week's topics and can tell you what's new. Memory is what turns a one-shot script into something that compounds. Hermes persists three kinds: declarative (facts the user told it), session (searchable past conversations via FTS5), and procedural (reusable skills). How to build a self-improving AI agent with skills, memory, and open models covers the strategies for working with each. How to organize your AI work like an operating system covers how the pieces fit on a single machine.
The three mistakes that kill free-agent jobs (and the cheap fixes)
These are the difference between an agent that works and an agent that wastes your afternoon.
Mistake 1 — The job is too vague
The agent has no idea what "clean" means, so it guesses, and it guesses wrong. Fix: write the job like you're writing it for someone on their first day. Name the file, the operation, the output format, and the save path. The job above ("weekly questions") is the corrected shape of a vague prompt.
Mistake 2 — Too many tools with no guardrails
An agent with 15 tools and a vague job picks tools at random. Fix: start narrow — three tools, get the job working, then add. Scope tools per profile when Hermes supports it. You will know you over-tooled the job when you come back 20 minutes later and the agent has called eight tools and produced nothing useful.
Mistake 3 — No checkpoint (the black box problem)
An agent that runs 20 minutes and then hands you a result is a black box. If it went wrong at minute two, you find out at minute 20. Fix: build in a stop. Have the agent tell you its plan before it starts, and have it write out what it found before it writes the final output.
In Hermes, you can do this with a two-turn job:
> Turn 1: Read the transcript at ~/calls/coaching_3.md and write out the
step-by-step outline you'd convert to a tutorial. Stop and show me the outline.
> [review the outline]
> Turn 2: Now write the full tutorial using that outline to ~/tutorials/.
Ten seconds of your attention saves ten minutes of rework. This pattern — review-then-write — is the single highest-leverage habit you can build into agent work.
What is the 20-minute job test (and why it matters when picking a free model)
Short tasks make every model look fine. The thing that separates a trustworthy model from an unreliable one is long tasks: does the model remember the original instruction from minute one at minute fifteen? Does it stay on task after eight tool calls? Does it notice when something didn't work and recover, or does it loop?
The 20-minute test is exactly what it sounds like: give the agent a job that you expect to take ~20 minutes of model work — read three documents, make 10 tool calls, and produce a structured output. Watch whether the agent:
- Remembers the original instruction from turn one to turn ten.
- Stays on task after multiple tool results come back.
- Notices when a tool call failed instead of pretending the failing result was the answer.
Free models vary wildly here. Hy3's combination of 256K context + explicit agent tuning is exactly what makes it pass this test where cheaper free models fail. For real-world routing logic across free and paid models, see our AI model task-routing guide for Qwen 3.8, Fable 5, GPT-5.6, and Kimi K3.
Why free changes how you build agents
When a model has a per-token meter, you build differently. You write shorter prompts. You give the agent fewer tools so it makes fewer calls. You stop it early. You don't let it explore. And a careful agent is a worse agent.
The whole power of an agent is that it can try, fail, notice it failed, and try again. That retry loop is where the magic is — but you hesitate to use it when every step costs money.
When the model is free, you stop hesitating:
- You let the agent run long.
- You let it retry.
- You give it 10 tools instead of three.
- You let it read the whole document instead of a summary.
This is why a free week or a persistent $0 model lane is the right time to test. Run the agent jobs you've been too cautious to run. A concrete one for builders: take every piece of content you've published — every post, every email, every description — and have a Hermes agent read it all and write down what your actual voice is, not what you think it is. That job needs long context. It needs to hold a lot at once. Normally you'd hesitate because it's a big run. On a free lane, just run it.
What this means for you
- If you're an individual builder or hobbyist: start with the Nous Portal Free tier + the bundled $0 Tencent Hy3. Set up Hermes on a $5/month VPS or your old laptop per the Hermes Agent setup guide. Build one tiny job end-to-end (the weekly questions job above is a good first one) before you scale up.
- If you run a small business: layer two free lanes for redundancy — Nous Portal for Hy3 (when available), local Ollama for offline/sensitive data, OpenRouter
:freeendpoints as a fallback. Read how to run a small business on AI agents for the deployment map. - If you're testing reliability before paid spend: use the free lane exactly as a testing window. Run the 20-minute job. Use the results to decide whether the model fits your business. When you spend real money, you'll spend it on a model you've already stress-tested at $0.
FAQ
Q: Is running Hermes Agent actually free, or is "free" a marketing hook?
A: The framework is MIT-licensed and costs $0 to self-host (source). The only thing that can cost money is the LLM endpoint behind it. You can run a free lane — Nous Portal Free tier ($0 models), OpenRouter :free endpoints, or local models via Ollama. If you stay inside those lanes, your bill is literally $0.
Q: Do I need a credit card to start? A: No. Nous Portal Free tier requires only an OAuth sign-in. OpenRouter free endpoints don't require a payment method during the free window. Ollama runs entirely on your hardware with no account. A credit card only becomes necessary if you move to paid models or bought-in Tool Gateway usage.
Q: Which free model should I pick for agent work? A: For multi-step long jobs, Tencent Hy3 is currently the best free option because it is explicitly tuned for tool calling and long context (256K), with vendor-reported scores of 78.0 on SWE-bench Verified and 90.4 on GPQA Diamond. For lightweight tasks (summaries, simple chats) any free OpenRouter model works. For offline / privacy-constrained work, run a local model via Ollama.
Q: Why do my agent jobs stall halfway through? A: Three usual causes — vague job instructions (Pillar 1), too many tools with no scope (Pillar 2), or a model that can't hold long context (the 20-minute test fails). Fix in that order. Write the job like you're onboarding a new hire; start with three tools; pick a model tuned for agent work.
Q: Will the Tencent Hy3 free window end and leave me stranded?
A: The OpenRouter tencent/hy3:free window ended ~July 21, 2026; after that it became a paid endpoint at $0.14/M input and $0.58/M output. Nous Portal currently lists Tencent Hy3 as a Free-tier model but free model lists change. Treat free lanes as testing windows — when a lane closes, you switch to another free model or pay the (cheap) per-token rate. Build agent jobs to be model-agnostic so swapping providers is one command.
Q: Can a free agent do real work, or only toy demos? A: Real work, within the rate limits. Common production patterns we've seen on free lanes: weekly question clustering, transcript-to-tutorial conversion, content-voice analysis, scheduled briefing digests. The constraint is not quality — it is rate limits and context ceiling. If your job needs 50 tool calls in five minutes, you'll trip OpenRouter's daily cap; schedule the job or move the heavy parts to a paid tier.
Discussion
0 comments