Hermes Agent v0.20.6 just shipped — and it's the biggest release window the project has ever produced. Released August 27 by teknium, this patch release rolled up ~525 merged PRs and ~1,313 commits across ~1,557 files into a single stable tagged release (v2026.8.27).
For context: between v0.20.5 (August 19) and v0.20.6 (August 27), the project merged 177,113 additions and removed 21,682 lines of code. That's more code churn in one week than many projects ship in a quarter.
Here's what actually matters for your daily agent workflow.
This is the headline feature. Previously, Hermes' browser tool launched a clean, incognito-like Chromium profile for every session. That meant no login cookies, no saved sessions, and your agent couldn't access Gmail, GitHub, or any site you were logged into.
v0.20.6 introduces consent-gated real-profile browsing. Your agent now launches in your default Chromium profile, giving it access to everything you're logged into. On Windows, a close-with-approval flow lets you control when the agent closes browser tabs — no more phantom tab closures on shared machines.
This single change unlocks a category of workflows: reading email, managing GitHub issues, posting on LinkedIn, editing Google Docs. The agent sees what you see.
The Hermes Desktop Browser finally got its own OS-level window, meaning it no longer shares space with your work applications. Combined with a managed SSH remote-update engine and a fleet profile rail, you can now:
The remote MCP catalog got a massive expansion. Hermes now ships with 50+ live-verified vendor-hosted MCP servers, including:
| Vendor | MCP Server |
|---|---|
| Cloudflare | DNS, Workers, R2 storage management |
| Grafana Cloud | Dashboard creation, alert rule management |
| Better Stack | Log monitoring, incident response |
| Railway | Project deployment, environment variable management |
| Postgres | Query building, schema inspection, data export |
| Kubernetes | Pod management, service discovery, rollout control |
These aren't theoretical — each server was live-verified against the vendor's actual API endpoints before shipping.
Context compression is a constant struggle for long-running agents. Hermes v0.20.6 switches to lean-tail compression as the default strategy. Instead of aggressively compressing the entire conversation, it now preserves the beginning and end of conversations (where users care most) and compresses the middle.
For a typical 100-turn conversation, this means your agent remembers the original task and the recent results — exactly what matters — while compressing 60+ intermediate turns that are less critical. In benchmarks, lean-tail compression reduced context window usage by ~40% on long-running agents while maintaining 95%+ task accuracy on the original instructions.
The tool_search tool now supports multi-query with stemming. When your agent needs to figure out which tool to use, it runs multiple search queries simultaneously and uses linguistic stemming to match variations of the same concept.
This reduces "tool confusion" errors — where the agent picks the wrong tool because it searched for "image generation" but the tool was registered as "image-gen" — by up to 40% according to internal benchmarks.
macOS users finally get opt-in OS-keychain encryption for stored secrets. No more per-launch Keychain prompts — your API keys and OAuth tokens are encrypted at rest using the native Keychain.
Three new models join Hermes' model picker:
| Model | Provider | Price | Use Case |
|---|---|---|---|
| GLM-5.3-Flash | ZeroOne/CoHere | Competitive | Fast reasoning, coding |
| MiniMax M3 | MiniMax | Free | General purpose |
| MiniMax H3 Max | MiniMax | Competitive | Video understanding |
The MiniMax M3 free model is particularly exciting — a capable general-purpose model with zero cost for users who sign up via the Hermes auth flow.
The release notes mention that "full curated release notes for this window will ship with v0.21.0." This is significant — v0.21.0 will document everything from v0.20.0 onward, including highlights, feature areas, and complete contributor credits. The v0.20.6 release is essentially the "raw" version of all these changes.
Expect v0.21.0 to be the most comprehensive release notes the project has ever published.
Pro Tip: To update, run hermes update from your existing install. For a fresh install: curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
Real-profile browsing alone makes v0.20.6 worth installing. Your agent can now read your email, manage your GitHub issues, post on LinkedIn — it sees what you see. Combined with 50+ remote MCP servers, lean-tail compression for long-running sessions, and the new free MiniMax M3 model, this release turns Hermes from a great tool into an indispensable one.
Real-profile browsing, 50+ remote MCP servers, lean-tail compression — and a brand-new free model. Run hermes update or grab a fresh install.