Quick comparison
Side-by-side feature snapshot. Scroll down for the full comparison table and narrative.
What changed in 2026
Claude.ai Cowork: a substantial platform upgrade
Claude.ai launched Cowork on January 12, 2026 for Max plan users on macOS, expanding to Pro on January 16, Team and Enterprise on January 23, and Windows on February 10. Scheduled Tasks arrived February 25, 2026. Dispatch launched March 17, 2026.
Cowork brings 50+ service connectors: Slack (Jan 26), Gmail, Google Calendar, Google Drive, Microsoft 365, Notion, Asana, Linear, Jira, Salesforce, and many more. Memory auto-generates from your chat history — no manual curation required. This is no longer a simple chat interface.
Scheduling: both capable, different constraints
Both platforms support recurring automated tasks, but with different constraints. Claude Cowork requires a Max or Pro subscription, runs on Anthropic infrastructure, enforces a minimum 1-hour interval, and your data leaves your hardware to execute. Hermes runs tasks at any interval (even seconds), on your own server, with full access to your memory and installed skills, with no subscription tier requirements beyond the model API you already use.
For most users scheduling hourly tasks, Cowork is more than sufficient. If you need sub-hourly automation, on-premises execution, or tasks that touch private infrastructure, the difference matters.
Shell access and execution environment
Claude Cowork executes code in a sandboxed isolated VM — appropriate for a multi-tenant hosted service, but it can't reach your local filesystem, environment variables, internal databases, or private network resources. Hermes runs scheduled tasks and agent tool use with full shell access: your filesystem, your tools, your env vars, your local databases, and your private APIs are all in scope.
Memory: automatic vs. inspectable
Claude.ai's memory system auto-generates facts from your chat history. You can view and delete memories, but the extraction process is largely opaque — you can't read the raw memory files or see exactly how decisions were made. Hermes stores memory as plain markdown files you can read, edit, or delete with any text editor. Eight optional external memory providers (vector DBs, graph stores, etc.) are available for more advanced retrieval. Nothing is a black box.
When Claude.ai is the better choice
Claude.ai wins decisively when you want zero server setup and the tightest possible integration with Anthropic's Claude models. Projects and Artifacts, computer use, 50+ managed Cowork connectors maintained by Anthropic, and a polished mobile app are all available without deploying anything. If data sovereignty isn't a requirement and you're comfortable with Anthropic's infrastructure, Cowork is a genuinely excellent platform.
Full comparison table
| Feature | Claude.ai Cowork | Hermes |
|---|---|---|
| Persistent memory | ✓ Auto-generated from chat | ✓ Layered, permanent |
| Memory inspectability | Limited (opaque) | ✓ Editable markdown files |
| Scheduled tasks | ✓ Cowork (since Feb 25 2026) | ✓ Self-hosted, always-on |
| Scheduling interval | Hourly minimum | ✓ Any interval (seconds+) |
| Shell / code execution | Sandboxed VM | ✓ Full shell, your server |
| File access | Sandboxed VM only | ✓ Full filesystem |
| Service connectors | ✓ 50+ (Slack, Gmail, Notion, Jira…) | Via agent tool use |
| Self-hosted | No (Anthropic infra) | ✓ Yes |
| Provider-agnostic | No (Claude models only) | ✓ Yes, any model |
| Data sovereignty | No (Anthropic servers) | ✓ Your hardware |
| Open source | No | ✓ MIT license |
| Mobile access | ✓ Claude app | ✓ Messaging apps |
| Always-on | ✓ Managed cloud | ✓ Permanent server process |
| Setup required | ✓ None (managed) | Server setup needed |
Who should use which
- Data sovereignty — everything stays on your hardware
- Any-interval scheduling — tasks every second, minute, or hour on your server
- Full shell access — reach local DBs, private repos, internal APIs
- Provider flexibility — swap Claude, GPT-5.4, Gemini, DeepSeek, or local models freely
- Inspectable memory — read and edit plain markdown files
- Open source — MIT license, audit and extend the code
- No subscription tier constraints — features not gated behind plan levels
- Zero server setup — managed infrastructure, no ops burden
- Tightest Anthropic integration — first access to Claude model improvements
- 50+ managed Cowork connectors — Slack, Gmail, Notion, Jira maintained by Anthropic
- Projects & Artifacts — polished document and project workflow
- Computer use — visual browser control and desktop automation
- Managed infrastructure — Anthropic handles uptime, scaling, security
Ready to run your own agent?
Hermes is open source, MIT licensed, and runs on hardware you own.