An assistant that remembers your work
Hermes brings memory, reusable skills, scheduled tasks and chat connections into an assistant you can host yourself.
Hermes Agent vs Orchestrator Zero
oz0 records agent jobs running on your infrastructure. Hermes centers on an assistant that remembers and reuses work across conversations. oz0 centers on operating agent jobs across machines, with shared execution controls and records. Both support multiple models and agent delegation.
What Hermes Agent is
Hermes Agent by Nous Research is an assistant you can run locally or host yourself. It combines persistent memory, reusable skills, scheduled tasks and specialist agents across a desktop app, terminal and messaging platforms.
github.com- Runs
- Self-hosted or managed
- Licence
- MIT
- Checked
- 30 October 2026
Reviewed 30 October 2026 against official documentation and release v0.22.5 (bf0203a). Ratings cover the open-source agent and documented optional integrations. Runtime behavior and benchmarks were not independently tested; hosted-service assurances are not assumed.
What do you want to operate?
Agent jobs across your machines
oz0 focuses on running jobs across a fleet, with shared permissions, approvals, model budgets and job records.
Several agents and machines
Hermes supports specialist bots, remote gateways and A2A peers. oz0 organizes execution around jobs and enrolled worker nodes.
Read the sourceCheck the execution path
Both have controls with limits. Hermes documents host rules and Docker sandbox options. oz0 checks supported calls through Edge; external harness tools can have different coverage.
Read the sourceSide by side.
Hermes Agent's ratings come from its own pages. A column's name opens its definition and how Orchestrator Zero does it. How we rated
Where it runs
Any machine you own
Yes
DocsYes: Your hardware or cloud
Run the agent and gateway on your laptop, server or VM; remote terminal backends can run tools elsewhere.
Linux, macOS and Windows
Yes: Native on all three
Native Linux, macOS and Windows installation is documented. Architecture and desktop-packaging limits are listed separately.
Self-hosted control plane
Yes
DocsYes: Local agent and gateway
The agent and gateway can run on your infrastructure. Nous Portal and Hermes Cloud are optional hosted services.
No inbound ports
Yes
DocsConditional: Depends on connection
Local use and polling-based chat need no public listener. Remote gateways, SSH backends and inbound A2A use reachable endpoints.
Fleet operations
Yes
DocsConditional: Gateway and peer management
Desktop manages several gateways and A2A routes tasks to peers. The full join, label, block and upgrade lifecycle for worker machines is not established.
Durability
Durable runs
Conditional: Recovery depends on path
Sessions and task-board state persist. Interrupted API runs can require reconciliation; this is not a general guarantee of continuing from the last completed step.
Governance
Tool policy outside the model
Conditional: Tool and backend coverage
Command deny rules, manual approvals and MCP filters constrain tools. Coverage varies by backend; the default smart approval mode also uses a model to assess risk.
Approvals in chat
Conditional: Chat prompts; timed waits
Risky commands can ask in chat. The reviewed gateway approval queue waits in process with a timeout; durable recovery of that pending call is not established.
Tool snapshots
Yes
DocsNo: Tool lists refresh directly
MCP tool-change notifications automatically refresh the registry. This is not a block pending approval of a changed description or schema.
Tenants built in
Yes
DocsConditional: Profiles and separate keys
Profiles separate configuration, memory and sessions; API gateways can have separate keys. Host profiles share normal user-level CLI state by default.
Cost
Budgets that stop a run
Conditional: Limits on costed MCP calls
Per-server MCP spending limits ask or refuse before costed calls. Totals reset on restart; these are not a complete model-spend ceiling for a job.
Security
Provider keys off the workers
Yes
DocsConditional: Docker sandbox scope
The optional egress proxy keeps real provider keys outside Docker tool sandboxes. The host agent still holds keys; other terminal backends are not covered.
Sandbox with an egress allowlist
Conditional: Docker; network setup needed
Docker isolation and allowlisted proxies are documented. Proxy settings alone do not stop direct sockets; enforce network restrictions separately.
Audit and compliance
Tamper-evident audit trail
Not documented
Stored sessions and tool records are documented. These sources do not establish a signed or hash-chained history of agent actions.
Erasure by key
Not documented
Session deletion and pruning are available. Per-person payload-key destruction with protection after backup restoration is not established.
Retention with legal hold
Conditional: Pruning; no documented hold
Configurable session retention and automatic pruning are documented. A legal hold preventing deletion is not established.
Openness
Model provider choice
Yes
DocsYes: Multiple vendors and local
Use multiple providers or local model endpoints, with configurable routing for delegated work.
MCP and A2A
Yes
DocsYes: MCP and A2A plugin
MCP tools are supported. The A2A plugin both calls peer agents and accepts tasks, with streaming and task management.
Where Hermes Agent is stronger.
Memory and reusable skills
The assistant searches past sessions and creates reusable skills from completed work, so later conversations can build on earlier ones.
github.comAn assistant in everyday chat
Desktop, terminal and messaging integrations provide direct ways to work with the same assistant and receive scheduled results.
github.comFlexible tool execution
Choose local, Docker, SSH or hosted terminal backends, and delegate tasks to specialist agents.
github.com
Differences to examine.
Compare the scope, conditions and editions in the source notes.
Find your first task.
Explore an example or read the public Docs preview. Software access is coming soon.