Hermes Agent Goes Global with i18n, Smart Skill Tiers, and Mac Sandbox: Platform Maturity Accelerates Past 135K Stars

Hermes Agent Goes Global with i18n, Smart Skill Tiers, and Mac Sandbox: Platform Maturity Accelerates Past 135K Stars
πŸ“‘ Table of Contents

Hermes Agent, the open-source AI runtime from Nous Research, has crossed 135,066 GitHub stars β€” up roughly 3,200 stars in just two days since the May 4 community salvage wave. But the star count is only part of the story. In the past 24 hours, three major platform features have landed or been proposed that signal a decisive shift: Hermes Agent is maturing from a powerful single-user CLI tool into a globally accessible, enterprise-ready agent platform. For ecosystem context, see our ultimate guide to agent frameworks.

🌐 i18n Support: Hermes Agent Speaks Your Language β€” PR #20231

On May 5, project lead Teknium1 merged pull request #20231 β€” a carefully scoped internationalization layer that adds display.language support for Chinese (zh), Japanese (ja), German (de), and Spanish (es).

What’s Translated

Rather than attempting a full, fragile i18n of every string in the codebase, the team took a thin-slice approach β€” targeting the ~19 highest-impact strings per locale:

Area What Gets Translated
CLI approval prompt Dangerous-command header, choice menu, prompt text, all outcome lines (allowed once/session/always, denied, cancelled, timeout)
Gateway replies Restart-drain notice, goal cleared / no active goal, approval expired, config read/save errors

What Stays English (Intentionally)

  • Agent responses (the model controls that β€” orthogonal to this layer)
  • Log lines, error tracebacks, debug output
  • Slash-command canonical names and descriptions
  • Tool outputs
  • Setup wizard, /help text, banner, spinner verbs

The architecture is clean and minimal: a new agent/i18n.py module with a catalog loader, t() helper, and language resolution that checks HERMES_LANGUAGE env > display.language config > en default. Translations live in locales/{en,zh,ja,de,es}.yaml, and the source English catalog is kept as the single source of truth. A full 21-test suite validates the layer.

β€œIf a user wants the agent itself to reply in another language, they say so in their prompt β€” orthogonal to this.” β€” PR #20231 description

This is a thoughtful, pragmatic choice. By decoupling the UI translation layer from the model’s language behavior, Hermes Agent can serve a global user base without compromising the flexibility of prompting. For more on the Hermes ecosystem, see our complete guide to AI agents.

🧠 Smart Skill Lifecycle Management: Auto-Tiering + Auto-Matching β€” PR #20644

Filed today by xyshanren, PR #20644 is a landmark contribution β€” it imports and generalizes features originally developed in the hermes-agent-cn Chinese community fork, now proposed for upstream integration.

The PR introduces 851 lines of new code across two complementary systems:

SkillTierManager (agent/skill_tier_manager.py)

Usage-based skill classification into three tiers:

  • Builtin β€” core skills, always injected into system prompt
  • Frequent β€” high-usage skills (max 10), auto-injected based on a rolling 7-day window
  • Archived β€” low-usage skills, zero token cost, loaded on demand

Promotion: β‰₯3 uses in 7 days β†’ promote to Frequent tier. Demotion: 7 consecutive idle days β†’ archive. Users can also manually pin, promote, or demote skills, and batch evaluation is supported.

SkillMatcher (agent/skill_matcher.py)

Auto-detects relevant skills from user messages using three strategies:

  1. Keyword exact match β€” matches trigger_keywords from SKILL.md frontmatter
  2. Context hint match β€” file extensions (.pdf β†’ pdf skill, .py β†’ python skill, etc.)
  3. Jaccard fuzzy match β€” on skill descriptions for semantic matching

The integration is clean: build_skills_system_prompt() splits skills into active/archived sections, periodic evaluate_promotions() runs every 20 tool iterations, and a new CLI subcommand hermes skills tier {status|pin|unpin|evaluate} gives users visibility and control.

This is the first major upstream contribution from the Chinese community fork β€” a powerful signal that Hermes Agent’s ecosystem is going truly global.

πŸ”’ Mac Local Sandboxed Agent Orchestration β€” PR #20643

Also filed today, PR #20643 adds sandboxed sub-agent orchestration for macOS. This lets Hermes Agent spawn, monitor, and kill isolated agent processes β€” Codex, Claude, OpenCode, and Pi β€” through a secure local-node interface.

Security features include:

  • Prompts delivered over stdin, not argv β€” task text does not appear in process listings
  • Shell-free argv β€” no shell injection vectors
  • Process groups β€” clean teardown of child processes
  • Bounded output β€” prevents memory exhaustion from runaway agents
  • Trusted workdir validation β€” denies unauthorized file-system access
  • Prompt delivery timeout β€” cleans up subprocesses before any handle is tracked

This is the latest in the Mac local-node series (stacked after PRs #20585, #20604, #20624, and #20639) and signals a clear direction: Hermes Agent is becoming a multi-agent orchestrator that can manage sandboxed child agents across your local machine.

πŸŽ™οΈ Bonus: STT Complete Hook β€” PR #20647

Rounding out today’s activity, PR #20647 adds an stt:complete hook event that fires after voice transcription completes. This lets user-space plugins in ~/.hermes/hooks/ react to voice transcriptions β€” for example, an echo plugin that sends the transcribed text back via Telegram. Context includes platform, chat_id, user_id, and the transcribed text.

πŸ“Š The Bigger Picture: Platform Maturation in 24 Hours

Metric Value Change
GitHub Stars 135,066 +3,224 since May 4
Open PRs merged (May 5-6) 10+ Including i18n, CI fixes, Feishu, kanban
New features proposed 4 Skill Lifecycle, Mac Sandbox, STT Hook, i18n
Active forks 20,634 +200+ since v0.12.0
Open issues 8,436 Community engagement remains high

What makes today noteworthy is not any single feature β€” it’s the pattern. The i18n layer, the community-fork contribution, the sandboxed orchestration, and the voice hooks are all moving in the same direction: Hermes Agent is shedding its early-stage tooling feel and acquiring the hallmarks of a serious platform β€” global accessibility, autonomous resource management, security isolation, and extensibility hooks.

The project is now averaging well over 100 commits per day and seeing contributions from the Chinese community forked ecosystem returning upstream. At this pace, v0.13.0 β€” likely dubbed the β€œGlobal” or β€œOrchestrator” release β€” may not be far off.


The Agent Report will keep tracking Hermes Agent’s blistering pace. Check back tomorrow for more.