01
30-second orientationBefore the table: what you're actually choosing
The short gave you the headline pick. The reason this guide exists is everything underneath it. These three agents differ most in places a benchmark never shows: where your memory lives, what one config line controls, and who sends you a bill when the agent runs all night. Read the table, then the three setup walkthroughs, then the trade-offs. By the end you'll know not just which to pick but exactly what you're signing up to operate.
- OpenClaw — open-source (MIT), Node/TypeScript. A gateway-first messaging assistant: one long-running Gateway process is the control plane that owns every session and channel, so it meets people on WhatsApp, Telegram, Slack, Discord, Signal, iMessage and more. Skills are
SKILL.mdfiles you author; community ones come from ClawHub. - Hermes Agent — open-source (MIT), by Nous Research. Built to run unattended and improve itself: its docs describe autonomous skill creation, a global
SOUL.mdidentity, layered memory on disk, cron scheduling, and self-improvement. Runs locally, on a VPS, in Docker, over SSH, or serverless. - Claude Code — Anthropic's agentic coding tool. Proprietary. The deepest single model and the most polished operator/coder copilot, but hosted and billed through a Claude subscription or the Anthropic Console (with Bedrock/Vertex as enterprise routes).
02
Side by sideThe head-to-head table
Every cell is attributed to the right project's own repo or docs. Read it as: no tool wins every row, and the rows you weight depend on whether you're running customer messaging, an unattended back-office agent, or a coding copilot.
| Axis | OpenClaw | Hermes Agent | Claude Code |
|---|---|---|---|
| Licence | Open-source (MIT) | Open-source (MIT) | Proprietary (Anthropic) |
| Runs on | Your devices / server; one Gateway process | Local, VPS, Docker, SSH, serverless — your infra | Anthropic-hosted; Console or Bedrock/Vertex for enterprise |
| Model access | Bring your own model + keys | OpenAI-compatible: OpenAI, OpenRouter, Kimi, MiniMax, GLM, custom endpoint | Claude by default; CLI/IDE can point at 3rd-party providers |
| Skills / extensibility | SKILL.md files you author; install from ClawHub | Auto-generated from your workflows + self-improving; Skills Hub (agentskills.io standard) | Skills/slash-commands you write; it also writes its own auto-memory |
| Memory + identity | Workspace memory; a SOUL.md persona (workspace-scoped) | MEMORY.md + USER.md on disk, sessions in SQLite; global SOUL.md | CLAUDE.md you write + auto-memory it writes, per project |
| Billing exposure | Whatever your model provider charges — direct, metered | Whatever your model provider charges — direct, metered | Subscription or Console pre-paid credits; agent use is metered at API rates |
| Best for | Customer messaging across every channel they already use | Unattended ops that learn on the job and keep memory + identity | Hands-on coding / operator copilot with the deepest model |
03
Setup walkthroughOpenClaw in practice — the gateway and channels
OpenClaw's whole model is the Gateway: a single long-running control-plane process that holds sessions and fans messages out to channels. You stand it up once, then bolt channels on.
- Run
openclaw onboard— the interactive flow sets up the Gateway, workspace, channels and skills in one pass. - Add a channel from the CLI, e.g.
openclaw channels add --channel telegram --token <bot-token>. Each channel gets its own block underchannels.<provider>in config. - Tighten who can talk to it: per-channel keys like
dmPolicy: "pairing"and anallowFromlist keep the gateway from answering strangers — set these before you connect anything customer-facing. - Author skills as
SKILL.mdfiles in your workspace (workspace / personal / shared / plugin scopes), and pull community skills from ClawHub — but review each one first (see the supply-chain note below).
OpenClaw's strength is reach (a dozen-plus channels from one process) and its risk is the same gateway being a single point you must secure. Lock down
dmPolicy and allowFrom on day one.04
Setup walkthroughHermes in practice — where memory and identity actually live
Hermes is the one that 'grows with you', and the reason that's more than a slogan is that its state is plain files on infra you own. You can open them, diff them, and back them up.
- Install with the one-line script from the Nous Research repo, then point it at any OpenAI-compatible provider — OpenAI, OpenRouter, Kimi/Moonshot, MiniMax, GLM, or your own endpoint.
- Identity lives in a global
SOUL.md— the agent's voice and defaults, consistent across every workspace and channel. - Memory is two markdown files in
~/.hermes/memories/:MEMORY.md(env facts, conventions, lessons it learned; ~2,200-char budget) andUSER.md(who you are / how you work; ~1,375-char budget). They're injected as a frozen snapshot at session start, so edits show up next session — handy to know when you tweak them by hand. - Conversation history is separate: every CLI and messaging session lands in SQLite at
~/.hermes/state.dbwith FTS5 full-text search, so it can recall an actual message from months ago — not a summary of it. - Skills it writes itself live in
~/.hermes/skills/(agentskills.io standard, installable via Skills Hub). Gate them:~/.hermes/config.yamlhas awrite_approvalsetting so the agent asks before it rewrites its own memory or skills.
Because memory, identity and skills are files you own, governance is a backup job, not a vendor-support ticket. Snapshot
~/.hermes/ and you can audit exactly what the agent taught itself.05
Setup walkthroughClaude Code in practice — providers and who bills you
Claude Code is the deepest model and the smoothest copilot, but it's the one where 'where it runs' and 'who bills you' are decisions you make up front — and they changed in 2026.
- Default path: a Claude subscription (Pro/Max) or the Anthropic Console. The Console route is pre-paid credits and auto-creates a 'Claude Code' workspace so you can see spend.
- To run it on another backend, set env vars:
CLAUDE_CODE_USE_BEDROCK=1(plusAWS_REGION) for Amazon Bedrock, orCLAUDE_CODE_USE_VERTEX=1withANTHROPIC_VERTEX_PROJECT_IDfor Google Vertex. Some surfaces also accept other third-party providers. - Project memory is
CLAUDE.md(instructions you write, checked into the repo) plus an auto-memory the tool maintains itself. Note a gotcha: some features — Anthropic's Session Memory among them — are gated to first-party API access and don't light up on every provider. - If you'll run it unattended, decide the billing path deliberately: a flat subscription is fine for hands-on coding, but agent-style continuous use is metered at API rates (next section).
Pin model versions on enterprise routes with
ANTHROPIC_DEFAULT_OPUS_MODEL / ..._SONNET_MODEL / ..._HAIKU_MODEL so a silent upstream model swap doesn't change your behaviour or cost overnight.06
Eyes openThe trade-offs nobody puts on the landing page
Each strength has a tax. This is the part the demos skip.
- OpenClaw's reach is also its blast radius. The same ClawHub registry that makes it powerful was the vector for an early-2026 supply-chain campaign — security researchers found malicious instructions hidden inside
SKILL.mdfiles, including payloads that pulled a macOS infostealer. The defence is boring and mandatory: pin versions, read every skill before install, don't auto-trust the registry. - Hermes accumulates state you now own. Self-improvement means it writes its own skills and curates its own memory — great, until you realise you are the backup, the reviewer, and the keeper of
SOUL.md. It's also more to stand up than a one-command toy, and you're responsible for the box it runs on. - Claude Code's risk is policy, not code. From 4 April 2026 Anthropic stopped flat-rate Claude Pro/Max plans from subsidising third-party agents and moved that usage to metered API-rate billing — a $200 Max plan had been running, by the company's own framing, $1,000–$5,000 of agent compute. It later reinstated third-party agent use on subscriptions, but with a catch: a fixed, non-rollover monthly credit (roughly $20–$200 by plan tier), then API rates beyond it. Some heavy users reported bills jumping up to ~50× (reported across multiple outlets). Brilliant tool — just don't wire an always-on business to one provider's pricing page.
07
The honest verdictPick X if…
Match the job to the tool. Most real setups run two of these for different layers, not one for everything.
- Pick Hermes if you want an agent that runs unattended, gets better at the jobs it repeats, and keeps memory + identity as files on infra you control — and you're comfortable owning the box and the backups.
- Pick OpenClaw if you want to meet customers on every messaging channel from one process, stand it up fast, and lean on a big skill library — and you'll vet what you install.
- Pick Claude Code if you want the sharpest hands-on coding/operator copilot with the deepest model, and you accept proprietary plus subscription/Console (or Bedrock/Vertex) billing.
- Whatever you pick: own the layer beneath it — your keys, your infra, and the brand/billing on top — so a vendor's pricing change is a config edit, not a business crisis.
08
Turning a build into a businessIf the agent is the product, who do you bill?
Two of these (OpenClaw, Hermes) are open-source and yours to run. The moment a client uses what you built, a new question shows up that none of the three answers: how do you put your name on it and charge for it? That's a layer above the agent. If you'd rather not build the white-label front, the multi-provider model access, and the metered billing yourself, that's exactly what Knotie's reseller platform is for — package the agent you assembled under your own brand and domain, point it at multiple model families through one OpenAI-compatible endpoint (swap the model with a one-line change, not a new integration), and meter usage per customer on built-in credit billing so you keep the margin. The agent does the work; you own the business around it.
This is optional plumbing, not a requirement for any of the three agents — you can absolutely DIY the brand + billing. Knotie just removes that build.
09
Run this firstYour 6-point pre-commit checklist
Before you build a business on any agent — open-source or not — answer these. Every 'no' is a future migration you're signing up for.
- Can I self-host it on infra I control, or does the heavy lifting run on the vendor's servers?
- Is the licence open (I can fork it) or proprietary (I'm a tenant)?
- Can I point it at my own keys / swap the model behind it without a rebuild?
- Does its memory + identity live in files I own and back up (e.g.
~/.hermes/,CLAUDE.md), not a vendor-only database? - If it has a skill marketplace, will I pin and review every skill — or auto-trust strangers' code?
- If the vendor changed pricing tomorrow, what breaks, and how many days to switch?
Get the next drop
New AI build guides + the occasional bonus template. No spam, unsubscribe anytime.
By submitting you agree to our Privacy Policy & Terms. Unsubscribe anytime.
You're in — check your inbox to confirm.
Frequently asked questions
OpenClaw vs Hermes — which is better for running an AI business?
Different shapes. OpenClaw is gateway-first: one long-running Gateway process meets customers across WhatsApp, Telegram, Slack, Discord, Signal, iMessage and more, with a large (hand-authored) skill library on ClawHub. Hermes Agent is built to run unattended and improve itself — autonomous skill creation, a global
SOUL.md, memory as files on disk (~/.hermes/memories/) with sessions in SQLite, and it runs on your own infra (local/VPS/Docker/SSH/serverless). Want broad messaging reach now → OpenClaw. Want an agent that learns on the job and keeps its memory + identity → Hermes.Where does Hermes actually store its memory and identity?
On disk, in plain files you own. Identity is a global
SOUL.md. Memory is two markdown files in ~/.hermes/memories/ — MEMORY.md (environment facts, conventions, lessons it learned) and USER.md (who you are, how you work) — injected as a frozen snapshot at session start. Full conversation history is separate, in SQLite at ~/.hermes/state.db with FTS5 full-text search, and skills it writes live in ~/.hermes/skills/. Back up ~/.hermes/ and you've backed up the whole agent.Can Claude Code run on something other than Anthropic's own billing?
Yes, with caveats. Beyond a Claude subscription or the Anthropic Console (pre-paid credits, auto 'Claude Code' workspace), you can point it at Amazon Bedrock (
CLAUDE_CODE_USE_BEDROCK=1 + AWS_REGION) or Google Vertex (CLAUDE_CODE_USE_VERTEX=1 + ANTHROPIC_VERTEX_PROJECT_ID). But it's still Anthropic's model and proprietary tool, and some features (e.g. Session Memory) are gated to first-party API access and won't appear on every provider.What changed with Anthropic's billing in April 2026, and does it still bite?
From 4 April 2026 Anthropic stopped flat-rate Claude Pro/Max plans from subsidising third-party agents (like OpenClaw) and moved that usage to metered API-rate billing — a $200 Max plan had been running, by Anthropic's framing, $1,000–$5,000 of agent compute. It later reinstated third-party agent use on subscriptions, but capped by a fixed, non-rollover monthly credit (about $20–$200 by tier), then API rates beyond it. Some heavy users reported bills up to ~50× (reported). It bites whenever you run Claude Code unattended on a subscription — plan the billing path deliberately.
What is SOUL.md, and which of these have it?
SOUL.md is a plain-text file that pins an agent's identity, voice and defaults so they persist across sessions. Hermes has a global SOUL.md (consistent across every workspace). OpenClaw also ships a SOUL.md persona, scoped to the workspace. Claude Code doesn't use SOUL.md — its equivalent is CLAUDE.md (instructions you write) plus an auto-memory it maintains itself.Are OpenClaw and Hermes open source? Is Claude Code?
OpenClaw and Hermes Agent are both open-source under the MIT licence, so you can self-host and fork them. Claude Code is proprietary — Anthropic's docs note it runs through a Claude subscription or the Console, with Bedrock/Vertex as enterprise routes.
What's the single biggest risk with each one?
OpenClaw: registry supply-chain — vet every ClawHub skill after the early-2026 campaign that hid malware in
SKILL.md files; lock down dmPolicy and allowFrom on the gateway. Hermes: you own the operations — backups, the box, and reviewing what it taught itself. Claude Code: pricing/policy — metered API-rate billing for unattended agent use. The hedge for all three: own your keys, your infra, and the brand/billing layer on top.Do I have to use all three, or pick one?
Most serious setups end up running two for different layers — e.g. Claude Code for hands-on building, Hermes or OpenClaw for the always-on customer-facing or back-office work. The checklist at the end of this guide is the test: self-hostable, open/forkable, swappable model, files-you-own memory, vetted skills, and a known switching cost. Pick per layer, not per brand.
Sources · OpenClaw — official GitHub repository (MIT, SKILL.md skills, Gateway control plane) · OpenClaw docs — Personal assistant setup (openclaw onboard) · OpenClaw docs — Channels config (channels.<provider>, dmPolicy, allowFrom) · Hermes Agent Documentation — Nous Research (SOUL.md, self-improving skills, providers) · Hermes Agent docs — Persistent Memory (MEMORY.md, USER.md, ~/.hermes/state.db, FTS5, config.yaml) · Hermes vs OpenClaw: Local AI Agents Compared — Turing Post (install, backends, ~/.hermes/skills, agentskills.io) · Claude Code Overview — Anthropic docs (subscription/Console, CLAUDE.md + auto memory) · Claude Code enterprise deployment — Bedrock/Vertex env vars, model pinning — Anthropic docs · Third-party agents lose access as Anthropic tightens Claude usage rules (Apr 4, 2026; $1,000–$5,000 framing) — PYMNTS · Anthropic reinstates OpenClaw and third-party agent usage on Claude subscriptions — with a catch (non-rollover credit) — VentureBeat · Anthropic blocks OpenClaw from Claude subscriptions in cost crackdown (~50x reported) — TNW · Malicious OpenClaw skills used to distribute Atomic macOS Stealer — Trend Micro · Hundreds of malicious skills found in OpenClaw's ClawHub — eSecurity Planet