The Hermes Context Diet: Slash Per-Turn Token Overhead the Official Way
Every turn re-sends your whole conversation to the model — and you pay for all of it. This practical guide walks the four layers Hermes offers to slim context: dual auto-compression, the v0.20 compaction overhaul (micro-compaction, N-user tail, ghost-skill defense, lean tail), per-model thresholds, and everyday hygiene with /context, /usage and /compress.
Read more
Hermes Compaction Gets a Lean Default: Big-Window Models Stop Hoarding 170K-Token Tails
compression.tail_mode now defaults to lean: compaction keeps a clamped 2.5%-of-window verbatim tail (10K–25K tokens) instead of the legacy formula's 100K–240K hoard on big-window models.
Read more
GLM-5.3-Flash Lands in Hermes: A 1M-Context Option in the Model Picker
z-ai/glm-5.3-flash joins Hermes model catalogs for OpenRouter (43 models) and Nous Portal (32 models): a 1,048,576-token context window at $0.075/M input and $0.25/M output — built for marathon sessions and big codebases.
Read more
An MCP Install Could Write HERMES_YOLO_MODE: Catalog Credentials Now Follow a Closed Schema
Before the fix, the MCP catalog install endpoint accepted arbitrary environment variables and persisted them to .env — a malicious or malformed catalog entry could flip security switches like HERMES_YOLO_MODE. Catalog credentials now follow a closed schema; undeclared names get HTTP 400.
Read more
hermes update Stops Sabotaging Your Work: Draining Gateways, Surviving Backends, Precise Windows Kills
Hermes' upgrade pipeline got a full overhaul: gateways drain gracefully over the control socket instead of being tree-killed, remote serve backends survive updates, the Windows updater only kills this install's hermes.exe, and image-managed installs refuse in-place updates.
Read more
Browser snapshots capped at 15,000 chars? Hermes makes the budget configurable — and kills the hidden LLM-summary cost
Hermes adds browser.snapshot_threshold (default 15000 chars, min 1000), and browser snapshots no longer pay for auxiliary-LLM summaries — oversized snapshots truncate-and-store, with the full tree kept on disk for read_file paging.
Read more
Running several Hermes Profiles means several Docker containers? One config key lets them share a single one
Hermes adds the terminal.docker_shared_container_key option: trusted profiles can opt into one shared persistent Docker container — set the same key and cut resource waste without merging environments.
Read more
Need Canva, Dropbox, or GitLab but stuck on OAuth? Hermes now installs 44 official remote MCPs in one command
Hermes' MCP catalog just grew by 44 official remote MCPs — Canva, Dropbox, GitLab, Strava, Cloudflare and more — installable with a single hermes mcp install command, with every tool surface audited and debloated.
Read more
Want Hermes to use a specific cloud sandbox as its terminal? Now you write a plugin — no waiting for core changes
Hermes makes terminal backends a pluggable subsystem: a third-party cloud sandbox can register itself as a terminal.backend value via a standalone plugin, no fork required.
Read more
Asked Hermes to search the same query four times? A 20-minute result cache stops repeat searches from re-billing the vendor
Hermes adds web.cache_enabled and web.cache_ttl_minutes: web_search and web_extract results are cached for 20 minutes, so parallel subagents searching the same query pay for one vendor call instead of N.
Read more
Hermes Natively Supports Lightpanda: 9x Faster, 16x Less Memory, with Automatic Chrome Fallback
Your agent lives on a small VPS, and every browser action spawns a 200MB Chrome that takes seconds to start. Hermes now ships native Lightpanda support: a Zig-built headless browser that is 9x faster, uses 16x less memory, and automatically falls back to Chrome when it hits something it cannot handle yet.
Read more
MCP Just Announced a New Roadmap — and Hermes Is Already on the New Spec
MCP published its new roadmap on August 22: a stateless core, Multi Round-Trip Requests, Tasks moving into the spec, agent identity with DPoP — five priority areas set the direction for the coming months. Hermes already runs mcp 2.0.0, which implements the 2026-07-28 revision, with MRTR, server/discover, and elicitation all in place.
Read more
Your Connection Looks Alive, But Nobody's Home: How Hermes Heartbeats Heal Silent Disconnects
Close your MacBook for lunch, switch Wi-Fi, or let your VPN reconnect — and Hermes keeps showing an interface that quietly stops responding. A freshly merged heartbeat mechanism lets the TUI and Desktop detect dead connections and rebuild them automatically.
Read more
I Ran Hermes Inside Obsidian for a Week: The Official Skill + Plugin Combo Made My Vault Feel Alive
A week after wiring Hermes Agent into my Obsidian vault, it stopped being a place that stores notes and started being a knowledge base that grows itself: the official built-in obsidian skill reads and writes notes, the community plugin turns the vault into the agent's working directory, and Smart Graph surfaces semantic links I never made explicit. Full week log, both official paths, and the pitfalls I hit.
Read more
Are Your Commands Running on the Wrong Server? Hermes Fixes Cross-Profile SSH Environment Leakage
If you run multiple Hermes profiles — one for the company server, one for your home NAS — watch out: after switching profiles, terminal commands could silently reuse the previous profile's cached SSH environment and run on the wrong remote host. PR #92156 (merged Aug 22) scopes the terminal environment cache per session, closing the leak.
Read more
Why Your Subscription Drained in Hours: Codex Context-Window Defaults, Explained
A bigger context window sounds better, but on ChatGPT Codex OAuth it silently burned through subscription usage. Hermes now defaults Codex slugs to the advertised 272K and turns the verified 900K window into an explicit -900k picker opt-in. Here is how context windows are decided, why compaction follows the window, and how to take control.
Read more