Hermes Agent v0.14.0 dropped on May 16, 2026 — nine days after v0.13.0. It is the largest release in the project's history: 633 merged PRs, 1,393 files changed, 165,061 insertions across 808 commits, 545 closed issues, 215 contributors in the release window. Upstream tag: v2026.5.16.
The release notes themselves run to thousands of words. Below are the 22 things you actually want to know, grouped by the question they answer. Every bullet cites the underlying PR(s) from the upstream notes so you can verify it.
---
Installation & distribution
1. pip install hermes-agent works from PyPI
After fourteen months of "curl this script," Hermes is a real PyPI package. The wheel ships the Ink TUI bundle and the shell launcher, so the full experience comes out of the box. pipx install hermes-agent works too. The old curl one-liner, Docker, and git-clone paths are all still supported — this just opens the door for the Python crowd. (#26593, salvage of #26350)
2. Debloating wave — lighter installs
Heavyweight backends (Slack / Matrix / Feishu / DingTalk adapters, hindsight client, codex app-server, Pixverse / Camofox / image-gen SDKs, voice/TTS providers) now lazy-install on first use. The [all] extras drop everything covered by lazy-deps, the installer falls back through tiers when a wheel doesn't fit your platform, and a supply-chain advisory checker scans every install for unsafe versions. Smaller footprint, faster install, fewer transitive vulnerabilities. (#24220, #24515, #25014, #25038, #25766, #21818)
3. Native Windows in early beta
A first-class native Windows path lands across CLI, gateway, TUI, and tools. PowerShell installer with auto-detected/auto-installed MinGit (~45 MB, no admin), Microsoft Store python stub detection, foreground Ctrl+C preservation. About 40 follow-up Windows-only fixes already merged. WSL2 remains the most battle-tested production path; the "early beta" label is doing real work. (#21561)
---
Models & providers
4. xAI Grok via SuperGrok OAuth — and grok-4.3 jumps to a 1M-token context
If you pay for SuperGrok, you can now use Grok inside Hermes by signing in with your xAI account — no API key, no separate billing. The wire-through also bumps grok-4.3 to a 1M-token context window, so you can drop whole codebases or research corpora into a single prompt. Includes proper handling for entitlement errors and an SSH-to-tunnel docs page for when you're SSH'd into a remote box and need to complete the OAuth flow. (#26534, #26664, #26644, #26592)
5. hermes proxy — your OAuth subscriptions, your localhost
Run hermes proxy and you get a local http://localhost:port endpoint that speaks the OpenAI API, backed by whichever OAuth provider you're signed into — Claude Pro, ChatGPT Pro, SuperGrok. Codex CLI, Aider, Cline, Continue — anything that expects an OpenAI-compatible endpoint — just works with your existing subscription, no API key required. One subscription, every tool. (#25969)
6. NovitaAI joins the provider lineup
AI-native cloud with a flat-rate Agent Sandbox, open-source models (Llama, Qwen, DeepSeek). Salvage of #7219. (@kshitijk4poor) (#25507)
7. OpenRouter Pareto Code router with min_coding_score
OpenRouter's Pareto router automatically picks the cheapest model that meets a quality bar. The new min_coding_score knob sets that bar for coding tasks specifically — Hermes routes to the most affordable model that's at least that good at code. Stop paying for top-tier models when a mid-tier one suffices. (#22838)
8. Codex app-server runtime for OpenAI / Codex models
Optional runtime that drives OpenAI's Codex CLI under the hood when you're on OpenAI or Codex paths. Session reuse, automatic retirement of wedged sessions, proper OAuth refresh classification — the plumbing that makes long agentic runs not fall over. (#24182, #25769)
9. Alibaba Cloud renamed to Qwen Cloud
Picker and config now use the brand the rest of the world uses. Existing config keys still work — no breaking changes — but the UI matches the actual brand. (#24835)
---
Messaging platforms — 22 total
10. Microsoft Teams — end-to-end
Teams launched as the first plugin-shipped platform in v0.12.0 (March 2026). v0.14.0 wires the full Microsoft Graph stack end-to-end: auth + client foundation (#21922), a webhook listener that receives Teams events (#21969), a pipeline plugin runtime (#22007), and outbound delivery via the existing adapter (#22024). Wire up the bot once, chat from any Teams channel, DM, or group. (Salvages of #21408–#21411.)
11. LINE — first-class via the LINE Messaging API
LINE dominates in Japan, Korea, and Taiwan. Hermes is now there natively. (#23197)
12. SimpleX Chat — decentralized, no user IDs
Privacy-focused messenger with no permanent identifiers. Hermes runs on it without exposing identifiers. (#26232 — salvages #2558)
13. Discord channel history backfill (default on)
When Hermes joins a channel or thread for the first time, it reads the recent message history before responding. No more "what are we talking about?" — the agent has the context that's already on screen for everyone else. (#25984)
14. Native button UI for clarify on Telegram and Discord
Multiple-choice prompts now show real platform-native buttons. Tap to answer — especially nice on mobile. (#24199, #25485)
---
Agent loop & sessions
15. /handoff actually transfers the session live
Mid-conversation, hand off to a different model, personality, or profile — every message, every tool call, every piece of context moves with you. Switch from a fast model to a deep-reasoning one mid-debugging, or pass a session between profiles for different parts of a task. (#23395)
16. /subgoal — append criteria to an active /goal
The Ralph-loop /goal keeps the agent running until a judge accepts the success criteria. /subgoal <text> layers extra criteria on mid-run without restarting the loop. (#25449)
17. Per-turn file-mutation verifier footer
After every turn that wrote or edited files, the agent receives a short footer summarizing exactly what changed on disk — paths, line counts, the actual delta. Catches its own mistakes when a write didn't land or got silently overwritten, instead of confidently claiming the function was added. (#24498)
18. LSP semantic diagnostics on every write
write_file and patch now run a real language server against the edited file and surface new errors back to the agent before the next turn. Type errors, undefined symbols, missing imports — caught immediately. Goes way beyond v0.13.0's basic Python/JSON/YAML/TOML linting because it's actual semantic analysis. (#24168, #25978)
---
Performance
19. ~19 seconds off hermes cold start
Skills cache + lazy Feishu adapter + deferred imports across the import graph + models.dev cache-first lookup + parallelized doctor checks + welcome-banner skip on chat -q. The hermes tools All-Platforms screen dropped from 14 seconds to under 1.5. (#22138, #22120, #22681, #22790, #22808, #22831, #22859, #22904, #22766, #25341)
20. 180× faster browser_console evaluations
Route through the supervisor's persistent Chrome DevTools Protocol WebSocket instead of opening a new session for each call. Pre-this, browser tool calls were the agent's single slowest operation. Post-this, they aren't. Real-world page interactions feel instant. (#23226)
21. Cross-session 1-hour Claude prompt cache
When you use Claude through Anthropic, OpenRouter, or Nous Portal, the prompt prefix (system prompt, skills, memory) caches for an hour across sessions. Start a /new session and the first response comes back faster and cheaper because the cache is still warm from your last session. Background memory review hits the cache too. (#23828, #25434, #24778)
---
Multimodal
22. vision_analyze returns pixels to vision-capable models
If the active model can actually see (GPT-5, Claude, Gemini, Grok-vision), vision_analyze now passes raw pixels straight to the model instead of converting to a text description first. You get the model's actual visual reasoning instead of a degraded text-summary round-trip. (#22955)
---
Bonus highlights
The 22 above are the structural beats. Twelve more worth a mention:
- •
x_search— first-class X (Twitter) search tool, OAuth-or-API-key. (#26763) - •
computer_usecua-driver backend — non-Anthropic vision models can now drive your desktop, with focus-safe operations, refreshed onhermes update. (#21967, #24063) - •Unified
video_generatewith pluggable provider backends — new video models drop in as one-file plugins. (#25126) - •Clickable URLs in any terminal — real OSC8 hyperlinks with hover highlight. (@OutThisLife, #25071, #24013)
- •Zed ACP Registry one-click install via
uvx, no npm. (#26079, #26120, #26234) - •Brave Search + DDGS as free web-search providers alongside Tavily / SearXNG / Exa. (#21337)
- •9 new optional skills — Hyperliquid, Yahoo Finance, api-testing (REST + GraphQL), unified EVM multi-chain, darwinian-evolver, osint-investigation, pinggy-tunnel, watchers (RSS / HTTP JSON / GitHub polling), and a full Notion overhaul for the May 2026 Developer Platform. (#23582, #23583, #23590, #25299, #26760, #26729, #26765, #21881, #26612)
- •
huggingface/skillsas a trusted default tap — community skills published there install with no extra config. (#26219, closes #2549) - •Sudo brute-force block + three dangerous-command bypass closures + tool error sanitization (no more prompt injection via error strings). (#23736, #26829, #26823)
- •
/sessionsslash command for browsing and resuming previous sessions. (@austinpickett, #20805) - •Plugin
ctx.llm+tool_overrideflag — plugins get first-class LLM access and can cleanly replace built-in tools. (#23194, #26759) - •API server exposes run approval events — long-running runs no longer silently hang at approval-required commands. (#21899)
---
Numbers from the release window
| Metric | Count |
|---|---|
| Merged PRs | 633 |
| Commits | 808 |
| Files changed | 1,393 |
| Insertions | 165,061 |
| Issues closed | 545 (12 P0, 50 P1) |
| Contributors | 215 |
| Tag | v2026.5.16 |
| Days since v0.13.0 | 9 |
Top community contributors (release window)
- •@kshitijk4poor — 38 PRs (Telegram cadence/streaming, security hardening, codex-runtime hygiene, NovitaAI provider)
- •@alt-glitch — 13 PRs (Markdown-table TUI,
HERMES_SESSION_ID, hindsight-client optional dep) - •@OutThisLife — 12 PRs (TUI turn segmentation, markdown link titles, OSC8 hyperlinks)
- •@austinpickett — 8 PRs (
/sessions, personality switching, cron modals) - •@helix4u — 5 PRs (Google Chat setup, Windows Ctrl+C)
- •@rob-maron — 4 PRs (Nous Portal as model metadata authority)
---