Nous Research · Alternatives
Hermes Agent alternatives: four reasons, four shortlists
People leave this tool for four structural reasons, not because something is better: prompts that can become training data by default, a sandbox that is off until you turn it on, a project that ships releases every few days, and a balance that drains in dollars rather than an allowance that resets. Each reason points somewhere different, so this page is four short lists rather than one ranking.
Why people actually leave this one
Your prompts are training data unless you find the switch
On the vendor's own gateway the policy permits inference payloads to be stored, used to train and fine-tune models, and shared with partners and model providers. The control is a single account-level toggle, and it limits certain payloads without saying which. For a private repository that is a hard conversation with a reviewer, and the usual escape — an enterprise agreement — is a bigger purchase than most readers of this page are making.
Nothing is isolated until you isolate it
The terminal backend ships as local, so an agent that decides to run a command runs it on your machine. Container backends exist, are documented in detail and are genuinely hardened, but each is a configuration step. What covers the default instead is an approval layer whose default mode is a second model judging the risk — which is clever, and is also not a boundary.
The project moves faster than your change process
Releases land every few days on a calendar version scheme, and the repository carries around thirty thousand open pull requests. For a hobby machine that is momentum. For a fleet with a review cycle it means the defaults you audited are not necessarily the defaults you are running, and the alternatives worth looking at are the ones with slower, dated release trains.
You wanted an allowance, not a balance
The gateway's credits are dollars, which makes the meter unusually honest and also unusually open-ended: heavy use spends faster, and the only brake is a spend cap you set yourself. Plenty of people prefer the opposite trade — a flat monthly subscription with an allowance that simply stops. That is a different product shape, and it is what most of the subscription agents sell.
If you only read one line.
opencode
The structural twin, and the closest thing to a like-for-like move. Another MIT client nobody sells, another vendor that makes its money on model access, and a published per-token card we have checked against our own snapshot. You keep the property that made this tool appealing and trade a broad assistant for a narrower coding one.
Claude Code
If the answer has to survive a security review and you would rather have an allowance than a balance. It is also the only tool on this list we have run ourselves, so our pages about it rest on sessions rather than on documentation alone.
OpenClaw
If what you actually wanted was the half that answers on Telegram and remembers things, and the coding agent was incidental. It is the other tool in this section built around a personal assistant rather than a repository.
Every alternative we checked.
Grouped by the reason for leaving, not ranked by quality. The first rows answer the free-client and metering questions, the middle rows answer the policy and isolation questions, and the last rows are here because a good share of people reading this wanted a different kind of agent altogether. We publish no score for these tools and are not implying one by the order.
An open-source terminal coding agent whose vendor also sells model access, at rates it publishes per token.
The same commercial shape without the breadth. The client is MIT and nothing sells it, so leaving does not mean starting to pay for something you had for nothing. The difference is what the vendor's own gateway looks like: a published per-token card rather than a dollar balance, and one we tested against our own snapshot across twenty-five models, where it held to the cent on twenty-three.
Tradeoff: it is a coding agent and only that — no messaging gateway, no cross-session memory of the same kind, no scheduler. If half of why you installed Hermes was the Telegram bot that remembers things, this replaces the other half.
Best for: keeping a free client while moving to a bill you can compute in advance.
The subscription terminal agent, sold inside consumer plans with an allowance rather than a balance.
The opposite trade on every axis that makes people leave. Spend is a subscription with an allowance that resets instead of a balance that drains; the commercial terms address the coding tool directly rather than covering it under a general platform policy; and it is the one tool in this section we have run ourselves, across dozens of published capability tests.
Tradeoff: a closed client from a single vendor, locked to that vendor's models, with an allowance the vendor sizes and can change. You are exchanging the ability to point the tool anywhere for the ability to forecast the bill.
Best for: a predictable monthly number, and terms written about the agent itself.
A personal assistant that lives in your chat apps and launches coding agents underneath when it needs one.
The nearest match for what Hermes is when it is not coding. Same idea of one agent reachable from everywhere with continuity between conversations, same open-source client, and the same bill shape: the software is free and the model is not. If the coding agent was the part you never used, this drops the part you never used.
Tradeoff: it delegates real coding work to other agents rather than doing it, so as a replacement for terminal work it is a wrapper around the thing you already had, with a second layer of configuration to get wrong.
Best for: the assistant half — memory, messaging, automation — without the repository half.
An Apache-2.0 terminal coding agent included on every tier of its vendor's consumer subscription, free one included.
The other free-software coding agent in this section, and the one that answers a different question: if you are already inside a consumer subscription, the agent is included rather than metered separately. The client is open and can be pointed at any OpenAI-compatible endpoint, so the escape hatch you rely on here exists there too.
Tradeoff: you trade a meter denominated in dollars for one denominated in nothing at all. The allowance is a shared weekly pool reported as a percentage, spent alongside chat, image and voice generation, and no tier states a quantity.
Best for: people who already pay that vendor and want the client to cost nothing extra.
OpenAI's coding agent, sold inside ChatGPT plans and metered in credits.
Worth knowing about precisely because Hermes can sign into it: the same subscription is reachable from either client, so this is one of the few moves on this page you can evaluate without changing what you pay. If the conclusion is that you want the vendor's own harness and support path, the switch costs nothing but configuration.
Tradeoff: the meter goes backwards. Credits are a private unit whose dollar price the vendor does not publish, so a bill you can currently read in dollars becomes one you have to infer.
Best for: teams already standardised on one vendor's plans and identity.
An open-source agent that lives inside the editor rather than the terminal, billed through your own keys.
The editor answer to the isolation complaint. Work happens inside the workspace you already have open, with changes reviewed in the editor's own diff view, which is a different kind of safety net from a container — weaker against a rogue shell command, much stronger against a silent edit. Inference is billed to your key, so the meter stays in dollars.
Tradeoff: it is tied to the editor. No messaging gateway, no scheduled runs, no headless mode for a server, and nothing to reach from a phone.
Best for: wanting the diff in front of you before it lands.
An editor with the agent built in, sold as a subscription with usage-based billing above the included amount.
The answer for people whose complaint is really about assembly. Nothing here is configured from a YAML file, there is no gateway to keep running, no backend to choose and no approval mode to reason about. For an organisation buying a tool rather than adopting a project, that is most of the value.
Tradeoff: closed source, and the billing shape people complain about most in this category — a subscription with usage-based charges above it, which is neither a clean allowance nor a clean balance.
Best for: a team that wants one product, one invoice and an administrator.
A small Apache-2.0 terminal tool that edits your repository and commits, with no gateway and no memory layer.
The minimalist end of the same idea. Free software, your key, your machine, no account anywhere — and a fraction of the moving parts, which is the direct answer to a project that ships every few days. If what you want is an agent that edits files and makes commits, this is that and nothing else.
Tradeoff: check how alive it is before committing. The repository's most recent push at the time of writing was four months old, and the front page still recommends models that have since been retired — a small surface is only an advantage while somebody is maintaining it.
Best for: the smallest surface that still does the job.
Pick by what you're doing.
A reviewer asked where the prompts go
The repository is private, the answer cannot be "there is a toggle in account settings", and an enterprise agreement is out of scope for a team of four.
Claude Code
The useful difference is not that one vendor is more trustworthy but that the commercial terms are written about the coding tool itself rather than about a platform that happens to include it. If you would rather stay on free software, the other move is inward: point this same client at a model on your own machine, where the question stops having an answer because nothing leaves.
Finance wants the number before the month, not after it
A balance denominated in dollars is honest but open-ended, and somebody upstairs wants a line item that does not move.
Claude Code
An allowance that resets converts the question from forecasting to budgeting. The cost is that the allowance is sized by the vendor in a unit you do not control — so this is a trade of one kind of uncertainty for another, and worth making only if the finance process cares about variance rather than total.
You want the free-client property and a rate you can compute
The training default and the balance both bother you, but paying for a closed tool bothers you more.
opencode
It keeps the part worth keeping — nobody sells you the client — and replaces a dollar balance with a per-token card we have independently checked against vendor list prices. The migration costs a configuration file rather than a licence, and you can run both for a week before deciding.
The agent is for you, not for your repository
In practice you use it to remember things, run errands on a schedule and answer from your phone, and the terminal half sits idle.
OpenClaw
Then the coding agent is the part you are maintaining for nothing. The assistant-shaped alternative does the half you use, with the same free-client economics, and hands coding work to a dedicated agent on the rare occasion it needs one.
Your change process cannot keep up with the release train
You need to pin a version, audit its defaults, and know that what you audited is what runs for the next quarter.
Aider
A smaller surface is easier to pin and read end to end, and there is far less that can change underneath you. Check the maintenance signal first — as of this writing the project has been quiet for four months — and treat that quiet as the price of the stability you are buying.
When staying put is the right call
Frequently asked.
What people search next.