Release info

Hermes v0.20.6 Rolls Up 525 PRs While DeepSeek Harness Ships 0.1.2-alpha.1: Another Self-Hosted Double Release Day

Another self-hosted double release day on Aug 27: Hermes Agent v0.20.6 rolls up 525 PRs (real-browser profiles, a 50+ vendor-verified MCP catalog, new models like GLM-5.3-Flash), while DeepSeek Harness v0.1.2-alpha.1 delivers a broad UX/multimodal pass, per-subagent model config and a pi-ai kernel bump to 0.84.2.

August 28, 2026 · Source: https://github.com/nousresearch/hermes-agent/releases/tag/v2026.8.27

Direct answer: On August 27 two core self-hosted projects shipped on the same day — Hermes Agent v0.20.6 rolls up ~525 PRs since v0.20.5 into a stable tag (headline: consent-gated real-browser profiles, a 50+ vendor-verified remote MCP catalog and new models), while DeepSeek Harness v0.1.2-alpha.1 (alpha channel) delivers a broad pass on session UX and multimodal, per-subagent model config, and bumps the pi-ai kernel to 0.84.2.

Hermes Agent v0.20.6: one big rollup

Officially a patch tag, but the volume is ~525 merged PRs / ~1,313 commits / 1,557 files changed (+177K/−21K), targeted at Docker images, hosted deployments and fresh installs. Highlights:

  • Consent-gated real-browser profiles: use your default Chromium profile for local browsing (Windows close-with-approval flow) — no more login-state islands;
  • Desktop browser gets its own OS window, plus a managed SSH remote-update engine and a fleet profile rail;
  • Big remote MCP catalog expansion: 50+ live-verified vendor-hosted servers (Cloudflare, Grafana Cloud, Better Stack, Railway, etc.);
  • TTL result caching for web_search/web_extract; lean-tail compression is now the default;
  • OS-keychain encryption for stored secrets (no more per-launch macOS Keychain prompts);
  • Updaters pause gateways over the control socket instead of tree-killing; image/package-managed installs refuse unsafe in-place updates;
  • New models in the pickers: GLM-5.3-Flash, MiniMax M3 (free), MiniMax H3 Max (video).

For architecture context see our deep dive and product page.

DeepSeek Harness v0.1.2-alpha.1: a big UX pass

A new version line after 0.1.1-rc (alpha channel, early-adopter caveat). Highlights from the official bilingual notes:

  • Performance & footprint: initial page load, session-loading transfer size and persistence storage all improved;
  • Session UX: auto-collapsing process traces, resizable width, per-turn fine-grained usage, turn navigation, code highlighting kept during streaming;
  • Multimodal: instant image echo, context compression counts image tokens, trace view shows image attachments, models can locate uploaded images;
  • Per-subagent model config: tasks can pin provider/model/reasoning effort (Claude Code and Codex providers supported);
  • Persistent-terminal fixes: macOS/Linux PowerShell startup misdetection, host stalls when Bash spawns many children, and more;
  • Kernel bump: pi-ai upgraded to 0.84.2 (the Pi 0.84.x line we covered last week) — more evidence of the harness ecosystem riding the Pi kernel;
  • Legacy APIProxy transport removed (use the Remote gateway with one-time token auth); Python Windows x64 artifacts added.

How to read it

Hermes packed half a quarter's work into a "patch" tag, focused on real-browser capability and MCP supply; DSH pours effort into session experience and subagent scheduling. Both point the same way: self-hosted agents are competing on "pleasant and controllable", not just "runs". For selection, cross-check the comparison hub and the Agent Harness entry in our glossary.

Sources

← 全部 News 本地部署教程 深度文章