Hermes Agent Inbox + Which AI Models To Pair With It: The Plugin Needs None, Your Jobs Need The Right One, With Real GoldieBench Scores

The Hermes Agent inbox runs no AI model at all, so the real model question is which models power the jobs it reports on and the agent you ask to read it.

The plugin is the inbox labs plugin from Nous Research, version 0.1.0: a read-only snapshot of failed cron jobs, stalled delegations, running work, waiting results, upcoming runs and open chats for one profile.

Below are the scores I read on the live GoldieBench models page on 8 November 2026, and how I would use them for each part of an inbox routine.

To plan which model runs where with my team, book a free SEO strategy session and bring your job list.

The Inbox Itself Needs No Model

The plugin README is clear: one CLI command, no model tools, no hooks, nothing running during chats.

The catalog entry adds that it opens its stores read-only and sends nothing anywhere.

So hermes inbox gives the same output whichever model you have configured.

Model choice only enters at three points: the jobs that appear in the view, the agent that interprets it, and whether you want a model involved at all.

What GoldieBench Scores Actually Measure

The methodology page sets the rules: every model gets the same fixed prompt, one shot, no iteration and no best-of-N.

Each result must be a single self-contained HTML file that runs in a browser.

Scores are Julian's own 0 to 10, and the leaderboard rank is the average across every task a model was scored on.

That is a building test, not a status-reading test.

No GoldieBench task measures how well a model summarises an inbox snapshot, so treat the scores as a general signal of capability, not a direct ranking for this job.

The Scores I Read On 8 November 2026

Every number above comes from goldiebench.com/models/ as read on that date.

Scores move as new tasks are added, so check the page before you commit.

Reading The Table For An Inbox Routine

The top four sit within half a point of each other, so for the jobs an inbox watches, the gap between them matters less than how you configure each job.

Fusion and Hermes MoA are both panels rather than single models, per their taglines, which suits a heavy weekly build better than a quick nightly tidy-up.

The three 1M-context entries, MiniMax M3, Kimi K3 and GLM-5.2, are worth a look for jobs that read a lot of material before they write.

None of that changes what the inbox shows: a failed job looks the same in Needs attention whichever model ran it.

Picking A Model For The Agent That Reads The Inbox

Per the README, the agent gets at the snapshot simply by running hermes inbox --json as a shell command when you ask.

Reading that JSON, checking the sources field and writing a short list of next steps is a light task.

My editorial pick is whatever model already runs your main chat, because any model near the top of this table can handle it.

Spend your strongest model on the jobs, not on the summary.

Picking Models For The Jobs That Show Up In It

This is where scores matter more.

A cron job that builds a weekly report page, or a delegation that drafts a long document, is closer to what GoldieBench tests.

The cron docs let each job carry its own model: set one with hermes cron create or edit using --model and --provider, or lock in the current main model with --pin.

A job can also pin its own reasoning effort with --reasoning-effort, from none up to ultra.

So a heavy build job can run a top scorer such as Claude Opus 5 at 8.27, while a routine check runs a lighter setting.

For long, multi-step agent jobs, Kimi K3's tagline on the page, tuned for long-horizon agent work, makes it a fair candidate at 7.89.

When one of those jobs fails, the inbox lists it under Needs attention, whichever model ran it.

For the job setup itself, see our Hermes Agent cron job guide.

When You Want No Model At All

A morning inbox report does not need reasoning.

Cron's no-agent mode runs a script on schedule and delivers its output as it is, with no tokens, no model and no provider fallback.

Put hermes inbox in a script inside the Hermes scripts folder and schedule it:

hermes cron create "every 1d at 07:30" \
  --no-agent \
  --script morning-inbox.sh \
  --deliver telegram \
  --name "morning-inbox"

Combining the two is my own wiring of documented features, so run it once by hand and check the message first.

A Small-Business Pairing Example

Take a small design studio running Hermes for client work.

1. A Friday cron job builds a client status page as one HTML file, pinned to Claude Opus 5 because building is what GoldieBench rewards.
2. A nightly job tidies the week's notes into a summary, set to a lower reasoning effort.
3. The morning inbox report runs as a no-agent script, so no model is involved.
4. If Needs attention shows a failure, the owner asks the main chat model to read hermes inbox --json and explain it.

A shop, a salon or an accountant can use the same split: strong model for building, light setting for routine, no model for the report.

What The Scores Cannot Tell You

• How a model handles pending approvals; the inbox cannot show those, because they stay in the chat process that asked.
• Whether a model will hit a dangerous command in cron; the security guide denies those by default under approvals.cron_mode.
• How the plugin itself will age: it is a labs plugin that Nous built but does not maintain, per the labs repository.

Email Is A Different Feature

Searching for mail handling? That lives in the Email gateway, which connects a separate mailbox set aside for the agent using IMAP and SMTP, limited to the senders you put in EMAIL_ALLOWED_USERS.

The model you pick for that gateway is a separate decision, and outbound jobs should say "draft only".

Where The Plugin Stands

It came from core PR #81507, was moved into the labs repo, and reached the catalog with Hermes Agent v0.23.0 on 5 November 2026.

Pick your job models from the table, keep the report model-free, and recheck both after each update.

Also On Our Network

• Agent Operatorsthe six-step setup for the read-only inbox, from install to a morning cron report
• AI Income Deskwhether the read-only inbox is worth it for owners who leave Hermes running overnight
• AI Tool Verdicta seven-criterion scorecard for the read-only inbox, and who should skip it
• agentos.guidefive prompts that turn the read-only inbox into a daily Agent OS routine
• aiprofitboardroom.comthe complete Hermes Agent Inbox guide on the AI Profit Boardroom blog

Pin your heaviest job to a top scorer, keep the morning report script-only, and book a free SEO strategy session or join the AI Profit Boardroom if you want help choosing models around the Hermes Agent inbox.