Key takeaways: Hermes is a serious open-source agent: persistent memory, self-improving skills, natural-language cron, subagents, 20+ messaging surfaces, dangerous-command approvals, and optional container isolation. Bob is the finished team product: managed Slack operations, 3,200+ app connections, per-person access, scoped memory with provenance, human approval before sensitive business writes, and a receipt on every completed run. Pick Hermes to own and shape the runtime. Pick Bob to put an accountable coworker to work today.
Two different operating models
Hermes and HeyBob get compared because both put an AI agent to work, but they sit on opposite sides of one question: who owns the operating layer? Hermes is Nous Research's MIT-licensed hermes-agent. You install, configure, secure, update, and monitor it. In return you get broad model choice, a deep skills and memory system, cron, subagents, rollback checkpoints, many messaging gateways, and several terminal backends. Its official security docs include smart/manual dangerous-command approvals, fail-closed timeouts, deny rules, protected paths, and hardened Docker execution. Calling that "whatever you build" would be unfair.
HeyBob is the opposite operating model. We run and test the runtime. Someone @mentions Bob in Slack, Bob works across 3,200+ apps in an isolated disposable computer, uses the asking person's access for personal systems, remembers relevant person/channel/company context, pauses with a readable approval card before sensitive writes, and posts a cost receipt at the end. The value is not that you could not assemble those pieces around Hermes. It is that your team does not have to.
Side by side
| HeyBob (Bob) | Hermes (hermes-agent) | |
|---|---|---|
| Who operates it | We do. Fully managed runtime; you operate the product. | You do. Self-hosted on your own infrastructure, from a $5 VPS to a GPU cluster. |
| License | Commercial product, hosted or self-hosted in your VPC on the Enterprise tier. | Open-source, MIT licensed. |
| Where work happens | Slack today, with Microsoft Teams in early access. | CLI plus 20+ messaging platforms, including Slack, Teams, Discord, Telegram, WhatsApp, Signal, and email. |
| Connections | 3,200+ app catalog plus API, MCP, and browser paths. | 60+ built-in tools plus MCP servers and whatever credentials you configure. |
| Before-write behavior | Sensitive business writes pause in Slack with the app and action named plus Approve, Always allow, and Deny. | Smart or manual dangerous-command approval, fail-closed timeout, permanent allow/deny rules, and protected write paths. Business-app approval behavior depends on the tool or MCP integration. |
| Accounting | Every completed run shows credits, duration, and tool-call count, with larger-job estimates and a hard spend cap. | Local execution and cron history are available; infrastructure, model billing, and any unified business receipt remain yours to operate. |
| Per-user connections | Personal apps default to per-user OAuth, so each query runs under the asker's role. | Credentials and OAuth behavior are configured by the operator; Hermes is not a managed per-employee connection catalog. |
| Memory | Person, channel, and workspace memory with provenance, correction receipts, and Slack edit/delete controls. | Strong persistent memory, user profiles, cross-session recall, and optional external memory providers. |
| Model | Internal tiered model routing; you can also mandate a model per workspace. | Works with any model you point it at. |
| Pricing | Plans scoped by monthly credit volume, unused credits roll over, hard spend cap. | Your infra bill plus your time. |
Pick Hermes if you want to own the runtime end to end, your team is comfortable hosting and operating agent infrastructure, and you value an open-source, MIT-licensed base you can read and change. It runs anywhere from a cheap VPS to a GPU cluster, works with any model, and brings its own memory, scheduler, and subagents out of the box.
Pick HeyBob if you want the team operating model already assembled: 3,200+ connections, per-person credentials, scoped memory with provenance, schedules and external triggers, Slack-native approvals, isolated execution, finished files, and a cost receipt on every completed run. We run the runtime so your team can delegate work instead of maintaining an agent platform.