The state of Hermes Agent — November 2026

TL;DR

The September report ended with Hermes Agent at 246,755 GitHub stars, 250 Atlas-tracked projects, and a thesis: Hermes was becoming a persistent society of agents you can talk to, schedule, steer, and inspect.[4]

The two named releases since then spent their headlines not on new agent abilities but on how people get into Hermes and where everything in it runs.

As of November 12, 2026:

• 262,917 stars on the core repo (+16,162 since September 18, +6.5%)[2][4]
• v0.23.1 is the latest tagged release; 13 release tags landed in this window, from v0.21.5 through v0.23.1[1][6][18]
• 318 Atlas-tracked projects (+70 like-for-like) with 1,494,338 combined GitHub stars (+192,394, +14.8% like-for-like, after two off-topic projects were removed)[2][4]
• Memory & Context was the biggest mover: 57 repos / 305,374 stars, up 19 repos and 27.1%[2][4]
• Skills & Skill Registries is still the largest category at 53 repos / 499,579 stars, but it grew by only five projects[2]
• v0.22 “Threshold” let a fresh desktop install set itself up on a free trial model with no API key, and added hosted-app connectors, one-click local models, tested SemVer stable releases, signed desktop packages (including the Microsoft Store), and a 600-entry reviewed plugin catalog[9]
• v0.23 “Hearth” made one gateway own every session on a machine, served the Desktop workspace in a browser tab (the Webapp, off by default), and grew the plugin catalog from about 600 to over 1,050 entries[17]
• Upstream changed how it ships: v0.21.6 was the first release cut by a new stable release pipeline, tags moved from calendar v2026.M.D to plain SemVer, and source installs now follow tagged releases by default[6][7][8][9]
• The new process misfired twice: v0.22.1 was built from the wrong commit, and v0.23.1 fixed a plugin catalog clients had silently ignored since November 9[11][18]

The shorthand: September was about continuity and coordination. November is about the front door and the hearth. Hermes is moving from a system you assemble toward a product you install, update, and keep living in, with more of the experience in upstream’s hands.

§1 — Eight weeks of delta

September figures are quoted from the prior report. November figures are the star counts stored in the Atlas data/repos.json as of November 12, 2026, refreshed by the scheduled rebuild rather than live-polled; core forks were not polled. v0.21.4 (tagged September 21, after September’s data cutoff) moves the release pointer but is not among the 13 window tags. On November 12 the Atlas removed TauricResearch/TradingAgents and jau123/MeiGen-AI-Design-MCP because neither references Hermes Agent. September baselines are restated without them, using their September 23 snapshot counts (108,308 and 1,772 stars), so every delta compares like with like.[2][4][19]

Three signals matter.

First: core growth more than halved its pace, and the ecosystem kept outpacing it. Core stars rose 6.5% in eight weeks, against 14.1% in September’s window. Atlas-tracked stars rose 14.8% on the restated baseline.[2][4] The core repo now accounts for about 17.6% of tracked Atlas stars, down from about 19.0%.

Second: star growth was concentrated in a handful of large projects. Measured against the Atlas snapshot of September 23, three non-core repositories added 78,481 stars: vectorize-io/hindsight, obra/superpowers and farion1231/cc-switch. That is 44.4% of the snapshot-to-snapshot increase.[2][3] None of the three describes itself as a Hermes project, so the Atlas total moves with the wider agent market as much as with Hermes.

Third: the count grew through curation, and new entrants were small. The 76 genuinely new projects carry 39,697 stars combined, about a fifth of the star increase. 49 of the 76 landed in three categories: Memory & Context (19), Plugins & Extensions (19), and Integrations & Bridges (11).[2][3] The Atlas is widening at the long tail faster than the long tail is gaining weight.

As in September, the Atlas measures compatibility and ecosystem surface, not direct Hermes adoption.

§2 — The release arc: Threshold → Hearth, and a new way of shipping

The September report described Quicksilver → Herald → Pantheon → Hardening. This window has two named releases and a change in the release machinery underneath both:

1. Threshold: get newcomers through the door.
2. Hearth: give every conversation one place to live.
3. Stable releases: make what users run a tested tag, not a moving branch.

v0.22 “Threshold” — the front door

v0.22.0 shipped on October 13. Its notes report, since v0.21.0, ~24,200 commits, ~6,080 merged PRs, ~7,180 issues closed and ~2,520 contributors, and call it “the largest window Hermes has shipped, about 2.5x the v0.21.0 cycle.”[9] That window starts at v0.21.0, so it overlaps the period September already covered.[9]

Nearly every headline concerns a first contact:

• First run: a fresh desktop install opens on a setup chat running on a free trial model (nous/welcome, once per OS user), so it needs no key. The CLI, TUI and gateway get no free model.[9]
• First app: connectors act in hosted apps such as Linear and Notion through the Nous tool gateway, creating a free connectors-only identity on first use. A Connectors page replaces the MCP tab.[9]
• First local model: Hermes installs and maintains llama.cpp, rates models against the machine’s GPU and RAM, and needs no account or key.[9]
• First install: sealed, signed desktop packages (a signed MSIX on Windows, including the Microsoft Store), plus a package manager (hermes pm) that switches environments only after a check passes.[9]
• First plugin: a catalog of 600 entries pinned to reviewed commits; eleven plugins that patched internals were delisted, and six memory providers moved out of core to “live with their makers.”[9]

Underneath, Hermes moved to one gateway and one desktop backend per machine for all profiles, and the codebase shrank “with zero behavior change” from about 1.06M to 698K lines (−34%).[9] The bill came in the upgrade notes: Python 3.14, MSIX-only Windows on 11 22H2 or newer, retired idle/daily session resets, and ended sessions inactive for 90 days deleted at startup by default.[9] The notes also list what was reverted, including parallel group-chat rounds in Bot Mode and full isolation of named profiles from the root auth.json.[9]

The patch tail — v0.22.1 to v0.22.7

Seven patch tags followed within three weeks. v0.22.1 (October 14) was built from the wrong commit: it shipped unreleased gateway work, and the Store app could stick on “Connecting.” v0.22.2 arrived the next day as “the release 0.22.1 was meant to be,” with a known issue: sessions that ran under the 0.22.1 gateway cannot be deleted.[10][11] The later patches rolled up between 13 and 776 merged PRs each; v0.22.5 made delegation pluggable, v0.22.6 first carried the Webapp, and v0.22.7 added lead routing for Group Chats.[12][13][14][15][16]

v0.23 “Hearth” — one place for every conversation

v0.23.0 shipped on November 5, reporting ~10,080 commits, ~1,950 merged PRs, ~850 issues closed and 666 contributors since v0.22.0.[17]

The central change is architectural. Before, the CLI, TUI, Desktop, editors, bots and cron could each run their own agent against the same chat, which the notes say “is how messages got written twice or went missing between surfaces.” Now one gateway per machine owns every session, and every surface attaches as a client: one queue, one history, one writer per chat.[17]

Most of the rest builds on that center:

• The Webapp serves the full Desktop workspace from the user’s own gateway, on the local network or behind a public URL with sign-in. It is off by default and single-user.[17]
• Every machine in one window: Desktop keeps all registered gateways connected, and a Group Chat can seat bots from different machines.[17]
• Group Chats that route themselves: a lead bot or automatic router picks who answers, cron jobs can post into rooms, and bot-to-bot exchanges stop after max_hops (default 4) past the last human message.[17]
• The catalog: opt-in plugin auto-update, a Verified tier for vendor-built plugins, and growth “from about 600 to over 1,050 entries.”[17]
• Guardrails: price tags and per-chat spend limits for paid MCP tools, and hermes curator lock for hand-curated skills. Built-in skills went from 58 to 44.[17]

The notes are candid about the migration cost. The one-gateway move landed in v0.22.0, and “this cycle fixed what that broke”: slash commands, /retry, /review, clarify cards, reasoning display, usage and ACP sessions.[17] Session auto-prune, on by default in v0.22.0, became opt-in again. The old browser tools were removed, and prompts that name them must be edited by hand.[9][17] One figure needs a caveat: the catalog section counts “~270 new listings,” which the notes do not reconcile with the 600 → 1,050 headline.[17]

The release process changed underneath

The least visible change in this window may matter most to anyone building on Hermes.

• Tags: v0.21.5 was tagged v2026.9.24; its successor was tagged v0.21.6, and every release since uses plain SemVer.[6][7]
• Pipeline: v0.21.6 (October 8) “is the first release cut by the new stable release pipeline: one attempt ref, a tested Docker image, and a receipt tag at publish.”[7]
• Channels: from v0.21.7, source installs of the official repository update to the latest published release instead of every commit on main; hermes update --set-channel main opts back in, and forks keep following main.[8] v0.23.0 made new installs start on stable too.[17]
• Gate: every release passes source CI, Docker, Nix, signed desktop builds and install/upgrade runs on every OS before publication.[9]
• Notes: patch releases now “roll up” a PR count, list new work as “undocumented here on purpose,” and defer curated notes to the next minor release.[7][18]

Fourteen days separate v0.21.5 and v0.21.6, which rolled up about 2,106 merged PRs.[6][7] Until this window, “Hermes” for a source install meant whatever was on main at the last update. Now it means a tested tag, which downstream projects can actually target.

The two misfires show the pipeline still catching up with itself. v0.22.1 was a build-input error the gate did not stop.[11] Then a November 9 site deploy pushed the published plugin-catalog.json past the 2 MiB cap clients accept. Every install silently fell back to its bundled catalog, missing new plugins, current pins and newer removed entries, until v0.23.1 raised the cap to 16 MiB.[18]

§3 — The strategic read: from agent society to installed product

The series has tracked one progression: harness → operating surface → agent society. This window adds a fourth state:

harness → operating surface → agent society → installed product

“Installed product” does not mean finished. It means Threshold and Hearth spent their headline space on the conditions of use: getting in, getting updates, finding extensions, and keeping one authoritative copy of each conversation. Four moves carry the argument.

1. Entry without prerequisites. Three new paths skip the API-key step: the nous/welcome trial model for desktop onboarding, the free connectors-only identity, and one-click local models.[9] Two run through Nous services. The third runs entirely on the user’s machine, which is the counterweight that matters most.

2. Distribution Hermes controls. Signed packages, the Microsoft Store, tested SemVer channels, one package manager, and a catalog with reviewed pins, delistings, version floors (requires_hermes) and a Verified tier.[9][17] The catalog is now a distribution channel with a review policy.

3. One center. September said the thesis works “only if those remain views into one coherent system.”[4] Hearth answers that at the architecture level: every surface, the Webapp included, is a client of one session owner.[17]

4. A smaller core. Memory providers, every messaging adapter, GitHub Copilot, mcp-serve, the agent importers, Singularity and Camofox moved into bundled, official or labs plugins. Built-in skills fell from 58 to 44, and source lines fell by a third.[9][17]

The boundary between “core” and “ecosystem” moved: former Hermes features are now plugins with named makers, and community work increasingly arrives as reviewed catalog entries. The risk moved too, from fragmentation to concentration. Connectors run through the Nous tool gateway, desktop onboarding runs on a Nous trial model, and the catalog is reviewed upstream, though off-switches exist (nous.guest: false, tools.connectors.enabled: false).[9] The migrations were opinionated too: removed browser tools, an OS floor raised twice in one window, and a session-deletion default flipped on and then off.[9][17]

Hermes became easier to enter and more opinionated to live in. The test for the next reports is whether those opinions stay reversible.

§4 — The ecosystem at 318 curated projects

September figures are quoted from the prior report; November figures are computed from the current Atlas data/repos.json. Against the September 23 snapshot there are 76 additions, 10 removals, and 5 renames or transfers: EKKOLearnAI/hermes-studio → EKKOLearnAI/ekko-studio, xintaofei/codeg → spacering-net/codeg, conorbronsdon/claude-context-os → conorbronsdon/agent-context-os, 335234131/agent-browser-mcp → Azheyc/agent-browser-mcp, and scavio-ai/hermes-agent → scavio-ai/scavio-hermes-skill. September counts for Multi-Agent & Orchestration (12 reported) and Plugins & Extensions (44 reported) are restated without the two projects removed on November 12. Three projects present in both snapshots changed category that day: two moved from Developer Tools to Memory & Context and one from Guides & Docs to Skills & Skill Registries. Removals reflect Atlas curation, not necessarily project shutdowns.[2][3][4]

Where the stars came from

The largest gains among projects present in both snapshots, excluding the core repo:

Per-project gains are measured against the Atlas snapshot of September 23 (commit 9a5fcc90f), the closest stored per-repo baseline; the September report did not print per-project figures for most of these.[2][3]

The largest newcomers are Orkas-AI/Orkas (9,291), sergebulaev/linkedin-skills (5,862), rlaope/oh-my-hermes (5,548), tutti-os/tutti (3,909) and agentscope-ai/ReMe (3,795); NousResearch/hermes-official-plugins also entered (667).[2] Two of the ten removals were dropped on November 12 for having no Hermes reference: TauricResearch/TradingAgents (108,308 stars on September 23) and jau123/MeiGen-AI-Design-MCP (1,772). The other eight total 2,831 stars at their September 23 counts, led by mohitagw15856/pm-claude-skills (1,393).[3]

Memory took over as the growth category

Memory & Context added 19 projects (two of them recategorised from Developer Tools) and 65,020 stars (+27.1%) and now holds 20.4% of tracked Atlas stars; vectorize-io/hindsight alone supplied roughly half the gain.[2][3][4] Most of its 19 newcomers came from one Atlas curation pass that screened sixteen memory integrations on November 9.

Upstream moved the same way: v0.22.0 moved six memory providers, Hindsight among them, out of core into catalog plugins and listed 65 memory entries in the new catalog, and v0.23.0 added more.[9][17] The timing is consistent with that move drawing attention to memory projects; the data does not show causation, and we do not claim it.

Skills slowed by count

Skills & Skill Registries is still the largest category by stars, but it grew by five projects, against fourteen in September’s window. Most of its +51,181 stars came from obra/superpowers and nexu-io/open-design.[2][3][4] Upstream, skills got smaller and more guarded: built-ins fell from 58 to 44, six vendor repos became trusted taps, and an in-tree org skill marketplace (Collective Wisdom V1) was merged and then removed.[9][17] The distribution energy in this window went to plugins.

Plugins: the Atlas count and the catalog count diverge

Plugins & Extensions grew from 43 to 62 tracked repos but totals only 5,266 stars, while the upstream catalog went from launch to over 1,050 entries.[2][9][17] The two measure different things. The Atlas counts GitHub repositories presenting themselves as Hermes projects; the catalog counts reviewed, pinned listings, many inside multi-plugin repositories. For plugins, the catalog is now the primary discovery surface, and the Atlas category is a partial view of it.

Multi-agent and integrations grew through new entrants

Multi-Agent & Orchestration is the category the TradingAgents removal changed most. Without it the category holds 32,275 stars, up from a restated 11,995, and nearly all of the gain came from three newcomers: Orkas-AI/Orkas, rlaope/oh-my-hermes and tutti-os/tutti.[2][3] Integrations & Bridges added ten projects but only 145 stars: a long tail of small bridges such as xP3ta/hermes-console, a native Android client for self-hosted Hermes.[2]

§5 — The big new axis: custody

July’s key word was evidence. September’s was continuity. November’s is custody: Hermes increasingly holds things on the user’s behalf, and decides who else may touch them.

Custody appears across both releases:

• Passwords: the browser agent signs in with logins from 1Password, Bitwarden or Hermes’s own vault, and “the model never sees the password.”[9] The vault key can live in the OS keychain, and each password-manager plugin touches only its own entries.[17]
• Tokens: Hermes attaches a plugin’s OAuth token only to the hosts that plugin declared, so the plugin never sees it.[17] Codex requests are made by core “so plugins never hold the token.”[8]
• Secrets per profile: an opt-in plugin host gives third-party plugins only their own profile’s secrets.[9]
• Money: a chat that would exceed its paid-tool spend limit asks first, even in YOLO mode.[17]
• The user’s own work: a locked skill cannot be changed by the agent through any tool, even under --yolo.[17]
• Installs and conversations: a failed update cannot half-apply, breaking plugin updates are held back, and each chat has one writer.[9][17]

Several of these rules hold “even in YOLO mode.” That is a design position: autonomy settings govern what the agent may do, custody rules govern what it may hold, and the user’s permissiveness does not dissolve the second set.

Custody also concentrates risk. One gateway is one point of failure, and v0.22.2 and v0.23.0 both spent effort on it.[11][17] A reviewed catalog is one supply chain, and when it outgrew the client cap, installs quietly used stale pins.[18] The more Hermes holds, the more its silent failures matter.

§6 — September predictions, scored honestly

1. State integrity becomes a release gate

Prediction: watch for migration tests, multi-writer stress tests, recovery drills and real-filesystem CI, and whether major releases spend less patch-window energy repairing state regressions.

Verdict: half right. A release gate arrived, but the failures moved. The gate is real: install/upgrade runs on every OS, E2E suites on release builds, the full suite on native Windows, and SSH/Docker backends against real daemons.[9] state.db gained lock timeouts that name the holder and repair commands for journal mode and crossed profiles, and the one-gateway design removes the multi-writer pattern at the root.[9][17]

But state-adjacent problems kept shipping. v0.22.1 came from the wrong commit and left sessions v0.22.2 cannot delete.[11] Auto-pruning old sessions went on by default and back off a release later.[9][17] The catalog was silently ignored from November 9 until v0.23.1, and still is on installs that have not updated.[18] No state.db campaign like v0.21.2’s appears in these notes, but that absence is weak evidence: patch notes in this window are deliberately terse.

2. Bot Mode proves—or fails to prove—the “agent society” thesis

Prediction: the proof would be durable workflows; the failure mode would be too many personalities, duplicated context and expensive conversations.

Verdict: the product deepened; real-workflow proof has no evidence either way. Group Chats gained lead-bot and automatic routing, room purposes and shared notes, cron delivery into rooms, bots from multiple machines, and (in v0.23.1) rooms that run on a gateway without Desktop. Parallel rounds were reverted; rooms run serially.[9][17][18] One September failure mode did surface: v0.23.0 says “bots no longer ping each other in endless loops” and caps exchanges at four hops.[17] Our sources contain no usage data on lasting multi-bot workflows.

3. Cron becomes the compounding layer

Prediction: watch for templates and public case studies showing recurring agents operating for weeks.

Verdict: the infrastructure shipped; the case studies have no evidence either way. v0.22.0 added self-healing schedules, one alert per incident, delivery that “needs positive evidence,” and webhook-triggered jobs.[9] v0.23.0 added a shared prompt cache per job, cancellation, per-job missed-run policy, and cron posts into Group Chats.[17] The notes mention one session store with about 35,000 cron runs: heavy use on one install, not a case study.[17]

4. Skill quality gets a measurable trust layer

Prediction: reproducible skill tests, sandboxed evaluations, provenance, attestations, compatibility and privacy-preserving success telemetry.

Verdict: not yet. Trust infrastructure arrived for plugins, not for skill outcomes. The plugin catalog gained reviewed pins, delistings, known-issue warnings, version floors, a Verified tier and scan findings on update.[9][17] Skills gained a lock, more trusted taps, a Skills Guard change (blocked installs in a 421-skill sample fell from 83 to 68), and opt-in telemetry on why installs fail.[9][17] None of this measures whether a skill improves results.

5. Interoperability moves from protocol support to cross-agent work

Prediction: watch for workflows where Hermes is one component in a heterogeneous system, and whether they stay debuggable.

Verdict: interop moved, but mostly inward. The most-shipped interoperability feature was migration into Hermes: first-time setup offers to import from OpenClaw, Claude Code, Codex CLI or Cursor, with importers added for Grok Bot and OpenBot.[15][17] Outward work was narrower: mcp-serve as an official plugin, experimental MCP events over webhooks, and A2A hardening against tokenless browser requests.[17] Cross-agent workspaces such as spacering-net/codeg keep appearing in the Atlas.[2] We found no evidence of heterogeneous workflows with shared identity and traceability.

§7 — What to watch next

1. Whether v0.24.0 delivers the record the patches deferred

v0.23.1 rolled up about 618 merged PRs and promised that “full curated release notes for this window ship with v0.24.0.”[18] With patch notes deliberately terse, the curated minor-release notes are the only complete public record. Watch whether v0.24.0 keeps the promise that nothing is skipped, and keeps its “Reverted in this window” section.

2. How much of the front door depends on Nous services

Onboarding, connectors and the free identity run through Nous infrastructure; local models are the fully local path, and onboarding offers a one-time local setup on capable machines.[9] Watch whether the local path stays equally prominent as the free tier and connectors grow.

3. The Webapp’s user model

The Webapp is off by default and single-user: “a signed-in tab acts as you.” Alongside it, v0.23.0 shipped an opt-in remote config backend for fleets with admin-locked keys.[17] Watch whether the Webapp stays a personal remote control or becomes a team surface, and what its security model looks like if it does.

4. Catalog governance at 1,000+ entries

The catalog now has tiers, delistings, version floors and auto-update, and it has already had one silent distribution failure.[17][18] Watch how removals reach installed copies, whether review keeps pace, and whether the Verified tier grows beyond its first five vendors.

5. Whether the one-gateway migration settles

v0.22.2 and v0.23.0 both carried gateway lifecycle fixes, and v0.23.1 added hermes gateway attach and gateway-hosted Group Chats.[11][17][18] If gateway fixes taper over the next releases, Hearth’s architecture held. If they keep dominating patch notes, the center is still moving.

About this report

Methodology: This report compares the September 2026 State of Hermes report with the current ksimback/hermes-ecosystem main-branch data and with the Atlas snapshot committed on September 23, 2026.[2][3][4]

The release arc uses the official notes for every tag from v0.21.5 through v0.23.1, with the most weight on v0.22.0 and v0.23.0.[6][9][17][18] This report treats v0.23.1 as the released boundary; post-v0.23.1 changes on main are not described as shipped.

Release-tracking disclosure: The Atlas’s own release tracking stalled from September 24 to November 12, 2026, because upstream changed its tag format from calendar versions to SemVer and our monitor did not recognize the new tags. This report was written after that backlog was ingested.

Data freshness: Star counts are those stored in data/repos.json on November 12, 2026, refreshed by the scheduled rebuild rather than live-polled; core forks were not polled. Per-project deltas use the September 23 snapshot (commit 9a5fcc90f, 252 entries). Release notes are the source for commit, PR, issue and contributor counts; approximate figures remain approximate.

Scope note: Atlas includes Hermes-native projects and larger adjacent projects with explicit Hermes compatibility. Combined ecosystem stars measure the reach of that compatibility surface, not unique Hermes users, installs, or exclusive adoption.

Companion to: The State of Hermes Agent — September 2026, July 2026, May 2026, and April 2026.

Not affiliated with: Nous Research. Hermes Atlas is an independent community project celebrating and mapping the Hermes ecosystem.

Corrections welcome: Open an issue.

Stay current

The Hermes Atlas newsletter covers the releases, projects, and patterns between full reports.

→ Subscribe at hermesatlas.com — free, 1–2× per month.

— Kevin · x.com/ksimback

Sources

• [1] https://raw.githubusercontent.com/ksimback/hermes-ecosystem/main/data/latest-release.json
• [2] https://raw.githubusercontent.com/ksimback/hermes-ecosystem/main/data/repos.json
• [3] https://raw.githubusercontent.com/ksimback/hermes-ecosystem/9a5fcc90fb3d9e2d9b78cd73855cf398985c2e0a/data/repos.json
• [4] https://hermesatlas.com/reports/state-of-hermes-september-2026
• [6] https://github.com/NousResearch/hermes-agent/releases/tag/v2026.9.24
• [7] https://github.com/NousResearch/hermes-agent/releases/tag/v0.21.6
• [8] https://github.com/NousResearch/hermes-agent/releases/tag/v0.21.7
• [9] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.0
• [10] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.1
• [11] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.2
• [12] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.3
• [13] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.4
• [14] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.5
• [15] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.6
• [16] https://github.com/NousResearch/hermes-agent/releases/tag/v0.22.7
• [17] https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.0
• [18] https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.1
• [19] https://github.com/NousResearch/hermes-agent/releases/tag/v2026.9.21