Hermes Release: v0.8.0

Hermes v0.8.0 shipped on April 8. It's a big one — 209 pull requests merged, 82 issues closed, contributions from 18 people. That's a lot to sift through, so here's the stuff that actually matters.

Background task auto-notifications

This is the headline change. Long-running tasks — training runs, test suites, deployments — now notify you the moment they finish, instead of waiting for you to ask "is it done yet?"

The old pattern was: kick off a task, leave the agent idle, come back later, ask for a status check. The new pattern is: kick it off, close the chat, go do something else, and the agent pings you on whichever platform you're on when the work is done.

If you run anything that takes more than a couple of minutes, this changes how you work with Hermes. It's the single biggest quality-of-life improvement in the release.

Switch models mid-conversation with /model

You can now change your provider and model without leaving the session. Type /model in any chat — CLI, Telegram, Discord, anywhere — and pick a different model. The conversation continues with the new brain.

This opens up a real cost/intelligence strategy: use a cheap fast model for boilerplate work, hit a hard problem, switch to Claude Sonnet or GPT-4 to solve it, switch back when the difficulty drops. Before v0.8.0 this meant restarting the session and losing context. Now it's one command.

Automatic fallback logic is baked in too, so if your chosen model errors out, Hermes drops to the next one instead of dying.

Idle-based timeouts

One of those fixes you don't notice until you realize your long task didn't die in the middle anymore.

Hermes used to time out based on wall-clock time — run too long, get killed. That made sense for stuck conversations but was a nightmare for anything doing real work. v0.8.0 replaces this with activity tracking. An idle agent times out; an agent that's actively working stays alive as long as it needs to.

For anyone running training loops, big builds, or multi-step workflows, this removes an entire category of frustration.

Native approval buttons on messaging platforms

Slack, Telegram, and Feishu now render actual buttons for approval prompts, instead of asking you to type /approve or /deny. Tap the button on your phone, approval happens, the session keeps going.

This sounds minor, and in terms of code it probably is. But it's a disproportionate improvement for anyone who runs Hermes as a gateway bot they interact with from their phone. Typing commands on a touchscreen is a tax. Buttons remove it.

Thread context gets preserved across approval exchanges too, so you don't lose your place in a multi-step workflow.

Centralized logging with hermes logs

Debugging used to mean digging through a few different places to figure out what the agent was doing. v0.8.0 consolidates all logs into ~/.hermes/logs/ — split into agent.log and errors.log — and adds a top-level command to tail them:

hermes logs

Pair that with the expanded hermes doctor diagnostics and you finally have a clear picture of what's happening inside the runtime without grep hunts.

MCP gets production-grade: OAuth 2.1 + malware scanning

Two serious upgrades to MCP support:

OAuth 2.1 with PKCE. Hermes now speaks standards-compliant OAuth when connecting to MCP servers. This matters for enterprise MCP integrations where servers sit behind real auth.

OSV malware scanning. When you install an MCP server package, Hermes automatically scans it against the OSV vulnerability database. If a package has known exploits, you get warned before it runs on your machine. Given that MCP servers run as subprocesses with whatever access you give them, this is the right default.

Together these move MCP from "useful but watch your hands" to "safe enough to point at a production integration."

100× speedup on large outputs

A nasty O(n²) regex backtracking bug got fixed, producing roughly a hundred-fold speedup on operations that touch big outputs. If you've ever had Hermes hang on a huge log file, a multi-megabyte diff, or a long-running command's output, this is the fix.

You won't see a changelog line that says "now fast" but you'll feel it the first time you throw a real-world file at it.

Smaller but notable

A handful of things that didn't get their own section but are worth knowing:

• Google AI Studio (Gemini) as a native provider. Direct Gemini access without going through OpenRouter. Context length is auto-detected via the models.dev registry.
• Matrix reached Tier 1. Full feature parity with Telegram and Discord — reactions, read receipts, encrypted media. Good news if you care about privacy.
• Free Xiaomi MiMo v2 Pro via Nous Portal. A vision-capable model for auxiliary tasks at no cost.
• New skills in the hub. p5js creative coding, manim for math animations, llm-wiki (Karpathy's LLM knowledge base), research-paper-writing, and a few others.
• Remote execute_code. The code execution tool now runs on Docker, SSH, and Modal backends, not just locally.
• Supermemory provider. A new memory backend with multi-container support for anyone who wants to try it.

The self-improving easter egg

Worth calling out one detail that embodies what Hermes is trying to be.

Hermes ran an automated benchmark of its own tool-calling behavior, identified five specific failure modes in how GPT and Codex models were handling tool calls, and then patched them itself. The release notes call it "self-optimized tool-use guidance." In practice, it means the agent looked at its own weaknesses and fixed them.

That's a strange sentence to type. It's also the kind of change that justifies the "self-improving AI agent" framing in the tagline. A lot of releases say that. v0.8.0 actually did it.

If you're upgrading

Nothing dramatic to worry about. Run:

hermes update

The updater handles config migration. If anything looks off afterwards, hermes doctor will tell you what to check. Bundled skills are synced during updates now, so you don't need to reinstall them.

Where to go from here

If you haven't used Hermes before, the learning path is the fastest way to get from zero to a working setup. If you just want the authoritative reference, the official docs are where to go. And if you want the full changelog with every PR, the release page on GitHub has it all.

v0.8.0 isn't a flashy release, but it's a substantial one. The seven changes above are the ones you'll actually notice in the first week of using it.