Hermes Agent 80% latency cuts and 51 updates, OpenClaw Mac app, and durable export tools
TL;DR
Hermes Agent rolled out dozens of stability, speed, and integration fixes across three days while OpenClaw added a Mac app and remote server catalog.
What shipped
Across this week Hermes Agent delivered repeated performance and reliability patches while OpenClaw released two supporting tools for local and remote use.
Hermes Agent
Nous Research shipped 51 updates to Hermes Agent between 22 and 26 July. The changes focused on faster cold starts, durable message delivery, secret management from password managers, and expanded model support including GPT-5.6 and Fireworks AI. Repeated daily releases addressed the same core issues with incremental refinements to the desktop app, CLI, and gateway.
- •Sessions export command Hermes sessions export now outputs Markdown, Quarto, HTML, prompt-only, and Hugging Face trace formats with age or workspace filters and optional secret redaction.
- •Provider and model additions Fireworks AI and DeepInfra became first-class providers with GPT-5.6, grok-4.5, Kimi K3, Claude Sonnet 5, and hy3 added to the catalog plus LM Studio JIT loading.
- •Multiplexed gateway routing A single bot token gateway can now direct guilds, channels, or threads to separate profiles that keep their own config, skills, memory, and secrets.
- •Durable delivery ledger Final responses are written to state.db before sending so a gateway crash triggers automatic redelivery on restart across Telegram, Discord, and Slack.
- •Subagent live transcripts Task delegation now streams human-readable logs that users can tail in real time while background completions survive restarts via a durable ledger.
- •Bitwarden and 1Password integration A SecretSource interface pulls secrets at load time from Bitwarden or 1Password vaults using op:// references with vault precedence and conflict warnings.
- •Smart command approvals An independent LLM reviewer judges flagged commands by default and user-defined deny rules block actions even in yolo mode with /deny feedback.
- •Subscription commands in TUI and CLI New /subscription and /topup commands let users check plans, preview costs, and apply changes inside the terminal or desktop billing tab.
- •Desktop rendering speedups Over 20 performance patches delivered 14x faster markdown lexing, virtualized diff panes, and reduced re-rendering under heavy loads.
- •First-turn latency reduction Cold-start time fell from 4.3 s to 0.9 s across CLI, gateway, TUI, and desktop with per-token painting and live reasoning streams.
- •Granular reasoning effort tiers New max and ultra tiers plus per-model overrides and session-scoped /reasoning controls let users tune depth.
- •Hermes Agent v2026.7.20 release The July 20 release shipped the cumulative set of export, routing, ledger, and performance fixes listed above.
OpenClaw
OpenClaw released two companion tools this week. A Mac app lets users run GPT-5.6 safely on local hardware while a remote catalog lists over 13,000 MCP servers, skills, and plugins for coding agents.
- •Plow Mac App Users can now run GPT-5.6 on OpenClaw and Hermes inside a dedicated Mac application that keeps data local.
- •Remote OpenClaw catalog A public index now surfaces more than 13,000 MCP servers, skills, and plugins that coding agents can load on demand.
What this means for you
For Vibe Builders: You can now export agent sessions in Markdown or Hugging Face formats and pull secrets from Bitwarden without writing code. The 80 percent latency drop and desktop speedups mean faster iteration when you test agent workflows. OpenClaw's Mac app gives a simple way to run the same models locally.
For Non-techies: Hermes Agent now recovers messages after crashes and lets you manage subscriptions inside the app instead of a browser. The Mac app from OpenClaw makes it easier to try frontier models on your own machine without extra setup.
For Developers: Durable SQLite ledgers, multiplexed gateways, and pluggable SecretSource interfaces reduce lost state in production agent deployments. Cold-start improvements and live subagent transcripts help when you benchmark reliability across restarts. Watch the MCP server catalog growth for new skills to integrate.
What to watch next
Track whether Hermes Agent v2026.7.27 adds more model providers or further ledger features. Watch OpenClaw for additional remote catalog updates or plugin examples that appear in the next release notes.
Harsh’s take
The volume of near-duplicate daily releases suggests the Hermes team is shipping incremental fixes rather than a single polished update. This pattern helps early testers but risks fragmenting attention for users who must decide which build to adopt. The practical action is to pin a specific release tag in production and test the delivery ledger against your own gateway crash scenarios before relying on automatic redelivery.
by Harsh Desai
Sources
- •[v2026.7.20](https://github.com/NousResearch/hermes-agent/releases/tag/v2026.7.20)
More AI news
- Daily RoundupMage-Flow-Edit-Turbo trends, Stable Audio Open ships, plus inbox cleanup tools
Microsoft and Owensong models hit Hugging Face trends while Replicate adds audio generation and PureBox.ai targets Gmail cleanup amid ongoing global model discussions.