By Cody Wise, Founder, Wise Media. Published 1 September 2026.
A self-hosted AI agent stack in 2026 has four decisions in it: which agent runtime, which model, which skills, and which server. For most businesses the answer is Hermes Agent on a 2 vCPU, 4 GB VPS, pointed at a hosted frontier model, with skills written as plain Markdown. That runs for a few dollars a month plus whatever your model calls cost.
Summary
- Agent runtime: Hermes Agent (Nous Research, MIT licensed) for a single operator or business. Buzz (Block, Apache-2.0) if you need humans and agents in shared channels, but it tells you itself that it is pre-1.0.
- Careful: several sites ranking above the project are run by third parties, not Nous Research. Some say so, some do not. Check the domain before you run the installer.
- Model: hosted frontier models still win on tool-calling reliability. Run local only if data residency or cost genuinely demands it.
- Skills: the instruction layer has genuinely converged. A Markdown skill file works across roughly forty products today. The execution layer has not converged at all.
- VPS: Hermes needs very little. Buzz needs meaningfully more, because it ships Postgres, Redis and an object store alongside the binary.
- Canada: Canadian data residency is a preference and a contract term, not a legal requirement for ordinary private businesses. The hosts selling one-click agent templates have no Canadian datacentre.
Table of contents
- What “self-hosted agent” actually means
- Best agent dashboard: Hermes Agent vs Buzz
- Best models to use in 2026
- Best skills, and whether they are portable
- Best VPS, and the Canadian question
- Is Canadian data residency actually required?
- What this costs
- Common mistakes
- Frequently asked questions
- The bottom line
What does “self-hosted AI agent” actually mean?
It means the agent runtime, its memory, its credentials and its logs live on a server you control, instead of inside somebody’s SaaS. The model itself may still be a hosted API. Those are two separate decisions and conflating them is the most common mistake in this category.
Three things change when you self-host. Your conversation history and business context stop being someone else’s training or retention question. Your API keys sit in your own environment. And the agent can reach internal systems that a hosted product cannot, because it is already inside your network.
What you take on in exchange is real: patching, backups, uptime, and key custody. If nobody in your business is going to own those, a hosted product is the honest answer and this article is not for you.
Best agent dashboard: Hermes Agent vs Buzz
First, check which site you are actually on
Search “Hermes Agent” and a good deal of the first page is not Nous Research. That is not automatically sinister, and the picture is mixed enough to be worth stating precisely rather than lumping together.
Some third-party sites are upfront. hermes-agent.ai carries a footer line reading “This is a fan website, not affiliated with Nous Research,” and links to the correct repository. Others are less clear. hermes-agent.org sets its HTML author metadata to “Nous Research,” describes the project as “By Nous Research” in its header, closes with “© 2026 Hermes Agent. All rights reserved.”, and carries no statement of non-affiliation anywhere on the page. Both link to the real GitHub repo, so neither is routing you somewhere malicious.
What they are not is authoritative. Several monetise by routing you to a paid hosted service, and the numbers do not always survive a check. hermes-agent.ai advertises “10k+ GitHub stars” for a repository that is comfortably north of 200,000, and its testimonial section quotes what appear to be its own internal build files, research/05-social-proof.md and src/data/comparisons.json, in the position where a source would normally go. Read them as marketing, not documentation.
The only official surfaces are:
github.com/NousResearch/hermes-agenthermes-agent.nousresearch.comnousresearch.com
This matters more than usual, because the standard install path is piping a shell script from a URL straight into bash. Copying that command off the wrong site is how you hand a stranger a root shell. Check the domain before you run anything.
Hermes Agent
Hermes Agent is an MIT-licensed, self-hosted agent runtime published by Nous Research and launched publicly in February 2026. It has since become one of the most-starred repositories on GitHub, comfortably over 200,000 stars as of September 2026. Note that this is a different product from the Hermes family of open-weight language models that Nous is better known for. The agent is deliberately model-agnostic and you can run it without ever touching a Hermes model.
What you get:
- A real web dashboard, served on port 9119 and bound to localhost by default, covering configuration, API keys, MCP servers, sessions, logs, analytics, cron jobs, memory, credentials and skills.
- An in-browser terminal that spawns the agent’s terminal interface behind a pseudo-terminal and renders it through xterm.js.
- A skills system where memory stores facts and skills store procedures. The agent writes its own skills after complex tasks and refines them with use.
- No telemetry. Per the official FAQ, API calls go only to the model provider you configure, and conversations, memory and skills stay in
~/.hermes/.
Official minimum is 1 CPU core and 1 GB of RAM, with 2 cores and 2 to 4 GB recommended. The bundled Docker Compose file sets a 4 GB memory limit, which is the number worth sizing against. It runs on Linux, macOS, WSL2, native Windows and Android via Termux, and there is an official Docker image.
Buzz
Buzz is a different shape of thing. Published by Block, Inc. and announced on 21 July 2026 under Apache-2.0, it is a self-hostable team workspace built as a Nostr relay, where channels, threads, DMs, voice, media, YAML workflows and git events are all signed events in one append-only log. Agents join as members with their own cryptographic keypairs rather than as bots with tokens.
Two accuracy notes. It is frequently reported as “Jack Dorsey’s Buzz”. Dorsey chairs Block, but he is not named in Block’s announcement, its engineering blog, or the repository. And the popular “Slack plus GitHub plus AI agents” line oversells the second half: Block says itself that the git forge is early.
The genuinely interesting idea is identity. Each agent signs its own work with its own key, under a narrowly scoped authorization signed by its owner, so authorization does not erase authorship and you can revoke one agent without touching the human. If you have ever tried to work out which teammate’s bot token did something, you will see the appeal.
It connects to external coding agents through the Agent Client Protocol, and Block names Claude Code, Codex and goose explicitly as working today.
Which one
| Hermes Agent | Buzz | |
|---|---|---|
| Publisher | Nous Research | Block, Inc. |
| Licence | MIT | Apache-2.0 |
| Shape | Personal or single-operator agent runtime | Team workspace where humans and agents share channels |
| Stack | Single runtime, SQLite-class local state | Rust binary plus Postgres 17, Redis 7, S3-compatible store |
| Published minimum spec | 1 core, 1 GB (4 GB compose limit) | None published |
| Maturity | Actively released, production install path | Explicitly pre-1.0. README: “Not finished.” |
| Best for | One operator automating their own work | A team that wants agents as visible participants |
Pick Hermes Agent if you are one person or a small business automating your own operations. It is the lower-risk choice, the lighter stack, and the faster path to something running.
Pick Buzz only if the shared-channel model is the actual point, and read its own maturity language first. Its README says “Not finished. We will tell you what works and what doesn’t”, its security policy says it is pre-1.0 with no long-term support branches, and the Windows installer filename literally contains the word alpha. Block is being unusually honest. Believe them.
One security detail worth knowing before you put anything regulated into Buzz: its audit log hash chain is tamper-evident, not tamper-resistant. Block documents this plainly. It detects accidental corruption and single-row edits, but an attacker with database write access can recompute the whole chain.
Best models to use in 2026
Agents fail differently from chatbots. A chatbot that produces a slightly worse paragraph is mildly annoying. An agent that produces malformed JSON in a tool call fails hard and stops. So the thing to optimise for is tool-calling reliability, not benchmark scores on reasoning.
Hosted models
| Model | Input, USD per million tokens | Output, USD per million | Use it for |
|---|---|---|---|
| Claude Opus 5 | 5.00 | 25.00 | Long-horizon agent work where a failed run is expensive |
| Claude Sonnet 5 | 2.00 | 10.00 | The default. Best speed to intelligence ratio for most loops |
| Claude Haiku 4.5 | 1.00 | 5.00 | High-volume, low-judgment steps |
| GPT-5.6 Terra | 2.00 | 12.00 | Balanced alternative, 1M context |
| GPT-5.6 Luna | 0.20 | 1.20 | Cheapest credible tool-calling option |
| Gemini 3.7 Flash | 0.75 | 3.75 | Cheap long-context work. Promotional rate through 2026-12-31, then 1.50 / 7.50 |
The trap nobody puts in a comparison table: per-token prices are not comparable across vendors, or even across generations of the same vendor. Anthropic documents that its newer models use a tokenizer producing roughly 30 percent more tokens for the same text. A model at half the headline price is not half the cost if it counts tokens differently. Test on your own workload before you decide anything on price.
Two smaller traps. OpenAI bills prompts over 272,000 tokens at double input and 1.5x output for the entire request, which is easy to trip in an agent loop that accumulates context. And Claude 4.7 and later return a 400 error if you set temperature or top_p to non-default values, which breaks older agent scaffolds outright.
Open-weight models, if you must run local
| Model | Size | Licence | Note |
|---|---|---|---|
| Hermes 4.3 36B | 36B dense | Apache-2.0 | Current Nous model. Not Hermes 4, which is superseded and carries a restrictive licence |
| Qwen3.8-27B | 27B dense | Apache-2.0 | Strong general local pick |
| DeepSeek-V4-Flash | 284B total, 13B active | MIT | The only genuinely permissive frontier-scale option |
| gpt-oss-120b | 117B total, 5.1B active | Apache-2.0 | Benchmarked by OpenAI at its shipped quantisation, so no quantisation tax to discount |
| Muse Glimmer 30B | ~30B dense | Apache-2.0 | Meta’s current open weights. Llama as a lineage is effectively over |
Three honest cautions, because this is where most guides overclaim.
- Nous publishes no tool-calling benchmarks for any Hermes model. The case for Hermes on tool calling rests on training design and on wide adoption of its tool-call format, not on measured scores. That is a reasonable case. It is not a benchmark.
- Do not trust any BFCL score you see quoted for a 2026 model. The leaderboard has not been updated since April 2026 and none of the current generation appears on it. Where a lab and an independent run have scored the same model, gaps of nearly twenty points have shown up.
- Quantisation breaks formatting before it breaks judgment. Published research on compressed models found 4-bit quantisation preserves tool use within a few percent but degrades real-world task accuracy by 10 to 15 percent, and degrades JSON-structured output notably faster than plain text. For an agent that is precisely the wrong failure: a malformed tool call is a hard stop.
Practical hardware floor for local: a 36B model at 4-bit is roughly 22 GB of weights, which fits a single 32 GB card. But KV cache is a real line item on top, not a footnote, and bandwidth rather than capacity decides whether the agent feels usable. A machine with lots of memory and slow memory bandwidth will load a big model and then crawl.
Aggregators
OpenRouter remains the default aggregator and, usefully, charges no markup on inference. It passes through provider list pricing and takes its fee when you buy credits instead, currently 5.5 percent on card with a minimum fee, or 5 percent flat in USDC. So a hundred dollars of inference costs about a hundred and five, with zero per-token markup.
Nous Portal, Nous Research’s own inference layer, runs Free at $0, Plus at $20, Super at $100 and Ultra at $200 per month in USD, with roughly ten percent bonus credits. Worth knowing: its per-token rates on third-party models are often below the vendor’s own list price. That contradicts the usual assumption that going direct is cheapest, so it is worth spot-checking your specific model rather than assuming.
Best skills, and are they actually portable?
A skill is a folder containing a SKILL.md file: a written procedure the agent loads when it hits a matching task. Memory stores facts, skills store how to do things. The format needs only two fields, a name and a description, and the body is ordinary Markdown.
The honest verdict on portability, which almost nobody states plainly:
- The instruction layer has converged. This is real. A Markdown skill file with a name and a description works today across roughly forty shipping products, including Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor, goose, Zed, Warp, Devin and Hermes Agent. Several of them scan each other’s directories for compatibility. Some vendors shipped support before the standard was even formally published.
- The execution layer has not converged. The moment a skill depends on tool permissions, bundled script execution, deterministic activation, or a version pin, portability breaks in host-specific ways you cannot see by reading the file. There is no conformance test, no version field, no packaging format and no lockfile in the standard. Hosts truncate and cap differently: one caps the skill catalogue by size, another by a percentage of context, another allows only one active skill at a time.
So: write a skill once and run it anywhere is true for prose and false for behaviour. Practical consequence, and this is the actionable bit: keep your skills as written procedures with as little host-specific machinery as possible, and they will genuinely travel. Reach for tool permissions and bundled scripts only where you must, and expect those skills to be tied to one runtime.
One governance note for anyone making a long-term bet. The tool protocol (MCP) and the agent-to-agent protocol have both been donated to a neutral foundation. The skills format has not; it remains stewarded by a single vendor, however openly licensed. That is not a reason to avoid it, given the adoption. It is a reason to keep your skills simple enough to move.
Best VPS for a self-hosted agent
Size against the workload, not the marketing. Hermes Agent’s official floor is 1 core and 1 GB, but its own Compose file sets a 4 GB limit, so 2 vCPU and 4 GB is the honest entry point. Buzz ships Postgres, Redis and an object store alongside the binary, so 4 vCPU and 8 GB, with real disk, is the realistic starting shape.
| Host | Hermes tier (2c/4GB) | Buzz tier (4c/8GB) | Currency | Canadian datacentre |
|---|---|---|---|---|
| OVHcloud (Canadian storefront) | VPS-1, from $6.20 | VPS-2, from $11.64 | CAD | Yes, Beauharnois QC and Toronto |
| Akamai / Linode | Linode 4 GB, $24 | Linode 8 GB, $48 | USD | Yes, Toronto |
| DigitalOcean | $24 | $48 | USD | Yes, Toronto |
| Vultr | High Perf, $24 | High Perf, $48 | USD | Yes, Toronto |
| Hetzner (EU) | CX23, €5.49 | CX33, €8.49 | EUR or USD | No |
| Hostinger | KVM 1, $6.49 intro / $11.99 renewal | KVM 2, $8.99 intro / $14.99 renewal | USD | No |
Four things to read carefully before you compare those numbers.
- OVHcloud’s Canadian storefront quotes in Canadian dollars. That is unusual and it matters. The catch is that those figures are twelve-month prepay rates; the no-commitment monthly and renewal prices are not published on the page.
- Hostinger’s headline is an introductory rate on a 24-month prepaid term. Renewals run roughly 1.85 to 2.2 times the intro price. To Hostinger’s credit, their own tutorial tells readers to budget using the renewal price rather than the intro rate, and also recommends a bigger plan than their Hermes landing page’s hero price implies. Take the tutorial’s advice over the landing page’s.
- Hetzner’s EU pricing is still very cheap; its US pricing is not. A June 2026 adjustment raised some US plans several-fold. Older roundups calling Hetzner universally cheapest are out of date.
- Akamai publishes flat pricing across regions, so Toronto costs the same as anywhere else. That makes it the cleanest “Canadian data, transparent price” option among the USD hosts. Vultr and DigitalOcean do not publish per-region tables, so a Toronto premium is possible and unverified.
The one-click templates, and the trade-off they hide
Hostinger and LumaDock both sell one-click Hermes Agent deploys. They work, and the convenience is real. Two things to know.
First, Nous Research does not endorse, list or partner with any VPS host anywhere on its own site. These are unilateral host products. What is being sold is convenience, not access, and the underlying install is a public one-liner that takes minutes.
Second, and this is the sharpest finding in this article: neither host has a Canadian datacentre. Today, the one-click convenience and Canadian data residency are mutually exclusive. If residency matters to you, you are installing it yourself on OVHcloud, Akamai Toronto, or a Canadian provider. That is a ten minute job, not a project.
Is Canadian data residency actually required?
For an ordinary Canadian private-sector business, no. It is a preference, a client contract term, or a procurement condition. It is not a legal requirement, and a great deal of content in this category says otherwise.
What the law actually requires under PIPEDA is accountability and comparable protection. An organisation stays responsible for personal information transferred to a third party for processing and must use contractual or other means to provide a comparable level of protection. The words “outside Canada” and “residency” do not appear. The federal privacy regulator has stated that PIPEDA does not prohibit transfers to another jurisdiction for processing, does not distinguish domestic from international transfers, and treats a transfer for processing as a use rather than a disclosure, so no fresh consent is triggered.
Notice is required. Consent is not. You must make plain that information may be processed in a foreign country. Once informed, individuals do not have an additional right to refuse. Put it in your privacy policy. Do not build a consent gate.
Where hard localisation genuinely exists, it does not reach ordinary businesses:
- British Columbia FIPPA section 30.1 was repealed in 2021. This is the most commonly repeated stale claim in Canadian hosting content, and it is still being published.
- Nova Scotia’s PIIDPA does require storage and access in Canada, and it reaches service providers of public bodies. If you serve Nova Scotia public sector clients, this is real.
- Federal government data direction is a procurement requirement that reaches you only as a contract term.
- Quebec’s Law 25 requires an assessment and a written agreement for transfers outside Quebec, not residency. Note the trigger is outside Quebec, so Montreal to Toronto counts.
- Alberta’s PIPA imposes a notification duty: name the country in your policies.
So Canadian hosting sells honestly on latency, trust, procurement eligibility and client-contract readiness. It does not need to sell on a legal requirement that does not exist for most buyers. This is general information, not legal advice, and nothing here asserts anything about your specific obligations.
What this costs
A realistic monthly total for a single-operator Hermes Agent setup has two lines: the server, and your model usage. The server is the small number. Using the Bank of Canada daily average for 31 August 2026 of 1 USD to 1.3866 CAD, a 24 USD per month VPS is roughly 33 CAD before your card’s foreign exchange spread, which typically adds another two to three percent. A Canadian-storefront VPS avoids that spread entirely.
Model usage is the variable that actually decides your bill, and it depends entirely on how chatty your loops are. Agents re-send context on every step, so a badly designed loop can cost many multiples of a well designed one running the same task. Measure a week of real usage before you extrapolate anything, and set a hard spend cap at the provider on day one.
Common mistakes
- Installing from an unofficial domain. The install path pipes a remote script into bash. Check the domain first.
- Confusing the runtime with the model. Self-hosting the agent does not mean self-hosting the model, and most people should not self-host the model.
- Comparing models on headline token price. Tokenizers differ by up to about thirty percent. Test on your workload.
- Exposing the dashboard to the internet. It binds to localhost by default for a reason. Reach it over SSH tunnel or a private network, not a public port.
- Putting a long-lived key in an environment variable and forgetting it. On both platforms the environment variable overrides safer key storage. That is how agents get identity, and it is also the obvious leak path.
- Quoting a benchmark that has not been updated in months. Especially BFCL.
- Budgeting on an introductory VPS rate. Renewal is often close to double.
- Assuming Canadian residency is legally required, then paying for it and still writing a privacy policy that does not disclose processing location. The disclosure is the part that is actually required.
Frequently asked questions
What is the best VPS for Hermes Agent?
For most people, a 2 vCPU, 4 GB plan. Official minimum is 1 core and 1 GB, but the bundled Docker configuration sets a 4 GB limit, so 4 GB is the sensible floor. If Canadian data residency matters, OVHcloud’s Canadian storefront quotes in Canadian dollars and has datacentres in Quebec and Toronto, and Akamai’s Toronto region prices the same as its other regions. If it does not matter, Hetzner’s EU plans remain the cheapest per unit of capacity.
Is Hermes Agent free?
The software is free and MIT licensed, with no telemetry. You pay for the server it runs on and for whatever model provider you point it at. If you run a local model on your own hardware, the inference itself is free too. Any site charging for access to Hermes Agent itself is selling convenience, not the software.
Hermes Agent or Buzz, which should I use?
Hermes Agent if you are one person or a small business automating your own work. Buzz if the point is a shared workspace where a team and its agents talk in the same channels, and you can tolerate pre-1.0 software. Block states plainly in the repository that Buzz is not finished, and its Windows installer is still labelled alpha, so treat it as promising rather than production.
Which model is best for AI agents in 2026?
For most agent work a mid-tier hosted frontier model is the right default, because agents fail hard on malformed tool calls and hosted models remain the most reliable at structured output. Reach for a top-tier model when a failed run is expensive, and a cheap one for high-volume, low-judgment steps. Run open weights locally only when data residency or volume genuinely justifies it, and expect structured output to be the first thing that degrades under quantisation.
Are agent skills portable between different AI agents?
Partly, and the distinction matters. A skill written as a plain Markdown procedure with a name and description works across roughly forty products today, which is a real and fast convergence. But anything depending on tool permissions, bundled scripts, deterministic activation or version pinning breaks in host-specific ways, because the standard has no conformance test, no version field and no packaging format. Keep skills simple and they travel.
Do I legally need to host my AI agent in Canada?
For an ordinary Canadian private-sector business, no. PIPEDA requires accountability and comparable protection for information transferred for processing, plus transparency about where processing happens. It does not require Canadian residency and does not distinguish domestic from international transfers. Hard localisation rules exist for Nova Scotia public bodies and their service providers and for federal government data, and Quebec’s Law 25 requires an assessment and written agreement rather than residency. Notice in your privacy policy is the obligation most businesses actually have. This is general information rather than legal advice.
The bottom line
Hermes Agent on a modest VPS, pointed at a hosted mid-tier model, with skills written as plain Markdown procedures, is the stack that works for most businesses in 2026. It is cheap, it is MIT licensed, it collects nothing, and it runs on hardware you choose, including hardware in Canada.
The two decisions people get wrong are taking install instructions from a third-party site without checking the domain, and paying for Canadian data residency they were told was a legal requirement when it is actually a business preference. Both are avoidable in about five minutes of checking.
Want this built and maintained for you?
We build and run agent infrastructure for Canadian businesses, including the unglamorous parts: key custody, patching, backups and a defined response window when something critical lands. Our growth and care packages cover ongoing operation, and our website and build packages cover the systems the agents plug into.
Tell us what you are trying to automate and we will come back with a scoped recommendation priced in Canadian dollars.
Related reading
- How to manage 10 or more client WordPress sites
- WordPress site hacked? The complete cleanup and hardening guide
- How much does website maintenance cost in Canada?
Sources
- Hermes Agent repository, Nous Research and official documentation and FAQ
- Buzz repository, Block, Inc. and Block engineering, “Buzz”, 21 July 2026
- Agent Skills specification
- OVHcloud Canada VPS pricing and Akamai cloud pricing, North America
- PIPEDA, Schedule 1, clause 4.1.3
- Office of the Privacy Commissioner, guidelines on cross-border processing
- Bank of Canada daily exchange rates
General information only, not legal, security or investment advice. Pricing and version numbers move quickly in this category; every figure here was read from a primary source on 1 September 2026 and should be verified before you commit budget.