Claude Cowork vs Hermes: open source keeps what the chat drops

Is Hermes a Claude Cowork alternative for a team?

For one person, yes. For a team, no. Hermes gives a single operator a capable agent on hardware they choose, and it does that part well. What it cannot do is hand a second person a change to read, because the work it finishes lives in a message thread and a database file on the host that ran it.

Two questions separate the cases: who owns the record of the work, and who reviews a change after the person who started it has moved on.

A gateway process is how Hermes reaches you

macOS and Linux take one curl line, Windows takes PowerShell, and Android installs from a Termux package repository (Hermes Agent on GitHub). hermes opens a terminal interface. hermes gateway puts the same agent behind Telegram, Discord, Slack, WhatsApp, Signal or email, and one memory follows you across every surface (the Hermes site).

Six built-in terminal backends carry the work: local, Docker, SSH, Modal, Daytona and Vercel Sandbox, and plugin backends add others such as Singularity and Apptainer. A $5 VPS is enough of a home. hermes model points the agent at Nous Portal, OpenRouter, OpenAI or a server you run. Releases arrive weekly or faster: v0.22.2 landed on 15 October 2026.

The defaults suit a personal tool. Messages arrive only from senders on an allowlist or through a paired DM. A shell command that looks destructive stops at one of three modes: smart, manual or off. A cron run or a webhook has nobody to ask, so it refuses (security). The person who installed the host holds the keys and the allowlist.

What a conversation keeps, and what it cannot

What a finished task leaves is a conversation: the command, its output, the paths it touched, the line that said the job was done. Scrollback in one app, plus rows in a SQLite database at ~/.hermes/state.db on the machine that ran the session, with full-text search over the messages (session storage).

Reading it back is easy. Reviewing it is a different act. Who signed this off? What moved since last month's version? Who reads it if the operator quits on Friday? The first is a scroll. The second has no answer in a thread, and the third ends where the laptop ends.

Claude Cowork makes another trade in a hosted shape. It sits beside Chat in the Claude app and runs a task through the folders and tools you choose. Inside a team, an admin sets access per group, caps spend and restricts tools. Approval is a plan setting: new tasks start in "Automatically approve" mode on Pro, Max and Team plans, and in "Manually approve" on Enterprise plans, while computer use, deletion and connector writes still ask (the Team and Enterprise article). A diff for a colleague to read afterwards is a separate artifact (Claude Cowork, checked October 2026).

The smallest record a company can hand over

Start with the agent itself. Each one is a markdown file under agents/ that holds the prompt alone, and its block in kortix.yaml names the description, the mode, the model and the grants before anything loads. One sentence below it says whose worker this is. Everything under that sentence is instruction, so a policy change arrives as an edit with a diff (the starter agent file).

Then the session. It checks the repository out onto a branch of its own, so a committed edit survives the machine and an uncommitted thought does not. What the next session should know sits in memory/, read before the first task begins.

Then the gate. Session work reaches main through a change request. Merge is default-deny for agents, unless an admin grants the merge capability. A person opens the request and reads the diff, which is where the question a thread cannot answer gets asked: does this match what we agreed? The reply keeps a name and a date beside it, and the next operator can open it a year later.

Channels close the circle without becoming the record. Slack, Microsoft Teams and email are live. Ask for the work from a chat window if that is where you are. What survives sits behind the window, in the repository.

Who each one is for

Kortix, when the work belongs to a team: three people rotate through one class of job, every seat needs the same instructions, and the approval has to be findable afterwards. One repository holds the setup, a reviewer reads the change, and the company keeps both. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

Hermes, when one person is the whole company, or a solo operator wants the assistant to follow them from app to app. The host is yours to patch, provider keys stay in your hands, and nothing sits between you and the model. A cheap VPS with the gateway up covers it, and an afternoon is enough.

Claude Cowork, when one deliverable is the whole job and installation is not on the table: the board memo assembled from a folder of meeting notes, the vendor shortlist, the security questionnaire answered against your own policy. State the goal, watch the steps arrive, take the file.

What each one costs to keep running

Kortix opens with a free project: $0, 500 credits for the month, no card. Model calls ride on a provider key you hold, or on a ChatGPT subscription you connect. Credits pay for the agent computer, and managed models need the Credits plan or Enterprise. Paid credits start at $50 a month (Kortix pricing, checked October 2026).

Installing Hermes is free. The bill comes from whichever provider you point it at. Nous Portal is optional: a free tier with free models only, then $20, $100 and $200 a month, each paid tier carrying credits worth a little more than its price (the Hermes site, checked October 2026).

Cowork is not sold separately. The bill is whichever Claude plan carries it: Pro at $17 a month, Max at $100, Ultra at $200, a Team seat at $20 (Claude Cowork, checked October 2026).