Should you try Hermes Agent?
Hermes Agent merits a restricted trial if you need persistent memory, reusable skills, scheduled work, and messaging access in one assistant. Our run on September 28, 2026, at commit 9a0a162 verified dependency installation only. A successful assistant conversation, provider connection, messaging delivery, and scheduled job remain untested.
Hermes Agent documents persistent memory, reusable skills, scheduled jobs, and a messaging gateway. The README at our measured commit describes session search, skills created after complex tasks, cron scheduling, and terminal or messaging entry points. Those capabilities explain the appeal of a persistent assistant, but they are documentation claims rather than results from our sandbox.
What our sandbox actually verified
Our sandbox completed the root npm installation; build and test were skipped. The attached record identifies our run as September 28, 2026, at commit 9a0a1625367242596d338ae2da541c4a1fc785a2. It specifies 3 CPUs, 8 GB of memory, and image lab-node:22.
The harness records command outcomes separately. Installation has a success flag, elapsed time, exit code, timeout status, and retained output tail. Build and test each have a skipped flag and the reason "no script". Installation returned exit code 0 in 111 seconds. That result establishes completion of the selected installation command, not completion of application setup or a workload.
The attached stored parser summaries report 1,209 dependency packages and 1,452 MB installed. They record 15 npm audit advisories: 11 high, 3 moderate, and 1 low. These are dated installation and advisory observations. They do not measure runtime speed, establish exhaustive package coverage, or prove application security.
The raw install footer disagrees with the stored dependency and advisory summaries. We have not reconciled that difference through complete logs or a new execution, so the stored figures remain the reported measurements. Audit scan coverage is unknown because the record contains no scannedPackages value. The footer cannot establish that coverage.
Skipped tests are not failed tests. The record detects Python tests, a tests directory, and CI workflows, but our selected root npm path did not run them. We also did not launch an assistant conversation, connect a provider, test messaging delivery, or exercise scheduled jobs.
MrKeyoor's testing methodology explains how command outcomes and measurements reach a review. Its current description provides provenance context; it does not establish that this older run used today's procedure or methodology version. This correction uses the supplied stored run without a new execution.
What permissions would you be granting?
Hermes command approval is not a complete containment boundary. The security guide says terminal commands run with the OS user's permissions and can reach paths that the dedicated file tools deny. It also describes Python operations inside execute_code that do not pass through terminal command-pattern checks; whole-script approval depends on the session context.
The guide separates messaging authorization from command approval. Allowing someone to contact the assistant and allowing the assistant to execute an operation are different decisions. Its documented controls include pairing, allowlists, command approval, file guards, credential filtering, and isolated execution backends. Our sandbox did not validate those controls.
For an initial trial, restrict the workspace and credentials, connect one provider, and verify one intended task. Test the relevant approval and isolation behavior before adding messaging access or unattended scheduling. This is an evaluation sequence, not a deployment outcome established by our run.
Which alternative fits your workflow?
OpenClaw is an alternative for an assistant spanning chat channels and device apps. Its README describes a gateway, messaging channels, companion apps, and device nodes. Consider it when those interfaces match your needs. Its README also says main-session tools run on the host unless sandboxing is configured, so choosing it still requires a containment decision.
gptme can expose its persistent shell, Python REPL, and file tools to an MCP client. Its README also describes a terminal agent, desktop and web interfaces, memory, and autonomous operation. Consider it as a replacement for a terminal workflow or as a complement when an existing MCP client needs those tools. The supplied evidence does not establish that it has a smaller installation or simpler deployment.
We did not install either alternative in this run. These choices follow their documented roles; they are not comparative measurements of reliability, speed, footprint, or security.
What the retained scores mean
Setup ease is 2/5 because installation completed while application setup and operation remained unverified. The rubric is: 1 for a blocked installation, 2 for completed installation with substantial setup or verification remaining, 3 for a verified basic workflow, 4 for verified routine setup, and 5 for a verified intended workflow requiring little intervention.
Documentation is 5/5 for breadth of signposted guidance. The README points to installation, configuration, security, integrations, architecture, and contribution documentation. On this rubric, 1 means basic installation guidance is missing, 3 means setup and routine use are covered, and 5 adds those operational and development topics; 2 and 4 indicate partial coverage between those levels. This score does not establish that every procedure works.
Community is 5/5 for documented development activity. The September 24, 2026 v0.21.5 release records 460 merged pull requests in its release window. The activity rubric assigns 1 when no dated activity is established, 3 when a dated release is established, and 5 when a release also documents substantial merged contribution activity; intermediate scores indicate partial evidence. Adoption, support quality, and response times were not measured.
Maturity is 3/5 because releases and documented operational controls exist while application validation remains incomplete. The rubric runs from 1 for no established operating or release structure, through 3 for documented structure with important validation gaps, to 5 for operational validation beyond this record. Scores 2 and 4 indicate partial evidence between those levels. The unresolved audit-summary discrepancy limits what this run can establish.