Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-09-11. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (MIT) + your model API costs | 2026-09-11 |
One-liner: Hermes v0.21.0 (released 2026-08-31, internally the “Pantheon” release) adds roughly 5,800 commits and ~2,100 closed issues on top of v0.20.0. A widely shared Chinese hands-on ranking of its 10 headline features is mostly right — with two claims that don’t hold up against primary sources, and one we couldn’t verify at all. This entry ranks the features by upgrade value and cites the release notes, docs, and repo API for every checkable claim.
Verified facts (checked 2026-09-11)
| Attribute | Value | Source |
|---|---|---|
| Release | v0.21.0 / v2026.8.31, published 2026-08-31 | GitHub Releases API |
| Scope since v0.20.0 | ~5,800 commits · ~2,475 merged PRs · ~2,100 issues closed | Release notes (same page) |
| License | MIT | Repo API (confirmed today) |
| Repo momentum | 244,492 stars, 50,641 forks; open issues + PRs combined ≈ 41.9k | Repo API, 2026-09-11 |
| Upstream | Nous Research (NousResearch/hermes-agent) |
GitHub |
Tier 1 — the ones that compound
Cron jobs that remember. Scheduled jobs now load and update persistent memory like any other agent; continuity=true carries each run’s output into the next, monitor-mode jobs skip the LLM when nothing changed, and cron output can land in a bot’s chat where it actually gets answered. This is the difference between a schedule and a pipeline that learns between runs. In the Chinese piece the author runs 13 tasks in three shifts and reports real incident receipts (a numbered failure with a cause, not just “it broke”) — that’s the selling point, and it matches what the release notes describe.
Profiles: one person, many agents. Each profile is a separate state directory — its own config.yaml, .env, memory, sessions, skills, and cron jobs — with shell aliases generated per profile. One profile per platform (a publishing bot for one channel, a research bot for another) with no shared state by default. The docs are explicit that two agent processes must never point at the same profile.
Browser, your real profile. The agent now drives the in-app browser directly — navigating, clicking, reading — and pages pop out to your system browser. The opt-in “real profile” mode runs a copy of your Chrome profile (auth files re-synced on session start) rather than remote-debugging your live one, sidestepping Chrome 136+’s block on the default profile directory. Default off. Keep it that way until you’ve read the docs.
Tier 2 — infrastructure you stop thinking about
Secrets: Bitwarden and 1Password, in-tree. Bitwarden Secrets Manager pulls API keys at process startup so nothing sits in plaintext .env; one machine-account token replaces N per-provider keys. 1Password is bundled as an optional skill. Note: Bitwarden support has been in-tree since v0.14+, so this is “mature,” not “new in 0.21” — the ranking is correct, the novelty framing isn’t.
hermes security audit. On-demand supply-chain scan against OSV.dev covering the Hermes venv, plugin requirement files, and pinned npx/uvx MCP servers, with --fail-on for CI. What it does not scan: globally installed packages or editor/browser extensions — the docs say so, and a real CI rule should.
The MCP command center. MCP servers and the catalog merged into one desktop page: paste-anything import, background health checks that nudge re-auth before a tool call fails, a cost/usage overlay, and hermes:// deep links that install servers behind explicit confirmation. The “catalog as bookmark folder” pattern is real; install only what a workflow needs.
Tier 3 — opt-in, priced in attention
Egress credential firewall (iron-proxy). The sandbox holds opaque proxy tokens; a local Apache-2.0 Go proxy terminates TLS and swaps in real credentials at the network boundary, pinned at v0.39 with an audit log. Default disabled. If your agent runs untrusted code against your own keys, this is the cheapest meaningful boundary in the release. Otherwise, skip.
Skills ecosystem. The release notes ship a productivity-skill wave and keep the self-improving loop at the core of the project. The specific “213 skills / 109 nodes” numbers circulating in the Chinese post are the author’s machine, not a product metric — treat them as one data point.
Unverified: “journey star map.” The Chinese post ranks a hermes journey command (skill graph, usage stats, animated playback) as one of the 10. We could not locate journey in the v0.21.0 release notes, the CLI command reference, or the current docs. It may be an optional skill or a very new command; until we confirm against primary sources, treat that ranking row as unverified.
Where the Chinese post got things wrong
- “iron-proxy v0.39.0” — correct, it is the pinned version, and the doc says “currently pinned.” Fine.
- “65+ top-level command families” — not a checkable number; the CLI reference lists a bounded set with subcommands. We dropped the claim.
- 44 keys in plaintext
.env— that’s the author’s own machine’s count, not a Hermes fact. Useful as motivation, not as a spec. - Bitwarden as “new” — in-tree since v0.14+; the release window made it prominent, not new.
Who should actually upgrade
Anyone running Hermes past a single interactive session: take the release as-is. Tier 1 lands immediately (cron memory and profiles are default-on behavior), Tier 2 wires in with one config change each, and Tier 3 costs you nothing unless you enable it. The release notes close with an honest framing — this is the version where a personal assistant starts to look like an operating system for agents — and for a solo operator, “one profile per job, keys in a vault, cron that remembers” is the three-line summary worth shipping.
Release notes, docs, and repo stats verified against primary sources on 2026-09-11. Feature ranking adapted from a five-day hands-on series by Agent工坊主 (WeChat, 2026-09-07 to 09-11); every claim above was re-checked against the release notes, CLI reference, and docs before publication. This entry follows our verification policy.