Comparison

Chickpea vs Hermes Agent

A self-improving personal agent with a good Slack gateway, or a workspace product built for Slack alone.

By Pejman Pour-Moezzi, creator of Chickpea.·Facts checked against the sources on September 4, 2026.

The short version

Hermes is the open source agent whose Slack support goes deepest. Native streaming, task cards, Block Kit, slash commands, a manifest generator, and a Bot Mode that builds a team of specialist bots with their own models and memory. It is also a personal agent at heart: it grows with one user, it lives on a box you keep up, and it gates access by a hand-maintained list of Slack member IDs.

Chickpea is the workspace product. Teammates are Slack handles anyone with the right role can create by asking. Accounts are connected per teammate and never reach the model. Consequential actions wait for a person. And the deploy is one button into your own Cloudflare account. Both are open source; they solve different problems.

Choose Hermes Agent if

You want one agent that learns your projects and follows you across Slack, Telegram, Discord, and WhatsApp.
You want the agent to write its own skills as it works and get better over time.
You want to run it on a $5 VPS, a GPU box, or a serverless backend, and you are happy operating it.
You want to use Nous models through Nous Portal, or any endpoint at all.

Choose Chickpea if

You want teammates for a whole workspace, managed by the people who use them, not by whoever has the shell.
You want per-teammate accounts, roles, and approvals as the default.
You want a deploy with nothing to keep running.
You want every member of the workspace to be able to talk to a teammate without being added to an allowlist.

Side by side

Twelve things that decide it.

Hermes Agent. Nous Research's open source, self-improving agent. One install, twenty chat platforms, a Slack gateway with streaming and task cards, and a Bot Mode for specialist bots. Chickpea. An open source Slack app that gives a workspace named AI teammates, on your own account. Here is how they line up on the questions teams ask first.

Chickpea12 of 12

Hermes Agent8 of 12, 4 partly

Yes Partly No

Criterion	Chickpea	Hermes Agent
Works inside Slack	Yes. Slack is the product: channels, threads, DMs.	Yes, as one of twenty platforms, through your own Slack app.
Several named teammates	Yes. Each teammate has its own handle, instructions, memory, and model.	Yes. Bot Mode builds specialist bots; the gateway takes several bot tokens.
Open source	Yes. Apache 2.0.	Yes. MIT.
Runs on your infrastructure	Yes. Your Cloudflare account, or Node on your own host.	Yes. Linux, macOS, WSL2, Docker, SSH, Daytona, or Modal.
Your own model keys	Yes. Anthropic, OpenAI, OpenRouter, or Workers AI with no key at all.	Yes. Nous Portal, OpenRouter, OpenAI, or any endpoint.
Choice of model provider	Yes. Four providers, pinned per teammate.	Yes. Any endpoint.
No seats, no markup	Yes. No seats, no credits. You pay the provider and Cloudflare directly.	Yes. Free software; you pay the host and the provider.
Managed from Slack	Yes. Create and change teammates by asking @Chickpea. Secrets finish in the browser.	Partly. Slash commands like /model in Slack; setup and config on the host.
Asks before consequential actions	Yes. Sending, deleting, publishing, and bulk changes wait for a person.	Partly. An allowlist of member IDs gates who can talk to it; no built-in gate on actions.
Accounts connected per teammate	Yes. 35 connector presets plus HTTP APIs and MCP servers, connected to one teammate each.	Partly. Per-bot model, memory, and skills; tool credentials live on the host.
Scheduled work	Yes. Recurring or one-time, created in Slack, posted to a channel.	Yes. Cron and routines.
Streaming, tables, live status	Yes. Streamed replies, native tables, task rows, live status, a signed footer.	Yes. Native streaming, task cards, Block Kit, slash commands.

Cost

What a 25-person team pays a month.

Three usage levels, from a few questions a week to a teammate in every thread. Chickpea's bill is the infrastructure and the model provider, paid directly. Open the numbers for every assumption.

The numbers and the assumptions

Usage	Requests a month	Chickpea	Hermes Agent on a small VPS
Light, 5 a person a week	541	$37	$44
Medium, 20 a person a week	2,165	$135	$142
Heavy, 50 a person a week	5,413	$330	$337

Chickpea

Cloudflare Workers Paid at $5 a month. Durable Objects, D1, and the cron trigger stay inside the plan's included usage at this volume.
Model usage at $0.06 a request on a mid-tier model, paid to the provider directly. Pick a frontier model for some teammates and that line goes up; pick a small one and it goes down.
Slack: named handles need a paid Slack plan, which a team of this size already has.

Hermes Agent

A small always-on VPS at about $12 a month. Nous suggests a $5 VPS is enough; a team gateway with several bots wants a little more.
Model usage at $0.06 a request, the same assumption as Chickpea, paid to the provider directly.
The chart is a near tie on purpose. Neither project charges anything. The differences are roles, approvals, credential isolation, and who operates the box.

Estimates from published prices on Nous Research's own pages and the sources at the bottom of this page. A request is one mention that reads a thread, maybe calls a tool, and answers. Your usage will differ; the shape of the bars is the point.

In detail

Where they differ.

The Slack experience

Hermes's Slack gateway is good. It generates the app manifest for you, runs on Socket Mode so no public URL is needed, streams replies through Slack's native streaming API, shows plan progress as task cards, formats with Block Kit, and answers slash commands. Tables past Slack's limits fall back to monospace, long output is capped at five posts, and custom status text needs an extra scope.

Chickpea does the same things and a few more, because Slack is all it does. Replies stream under the teammate's own name, task rows tick off above the answer, the status line under the composer narrates each step, seven or more rows come back as a native sortable table, and every reply ends with a footer naming the teammate and model with a link to change either. Teammates are Slack user groups, so @support is a handle people search for, not a bot user.

A personal agent that grows, or a workspace with roles

Hermes is built to grow with you. It curates its own memory, writes new skills when it solves a problem, and improves them while it works. Bot Mode turns that into a team of specialist bots with their own model, memory, and skills, coordinated through mentions in group chats. Access is gated by a list of Slack member IDs in the config; without it, the gateway denies everyone.

Chickpea starts from the workspace. Every full member of the Slack workspace is provisioned the first time they talk to a teammate; guests and Slack Connect users are not. Members create teammates and edit the ones whose policy allows it. Admins and Owners manage providers. Changes that widen a teammate's reach come back as a proposal to approve in the thread. Memory is one per teammate, readable and editable by you, and skills come from GitHub or skills.sh rather than being written by the agent.

Where the credentials live

Hermes runs tools on the host it lives on, with API keys in its environment. A bot with the Gmail credential has it because the process does.

A Chickpea teammate has the accounts connected to it and no other. The model picks an account by name; the secret is attached at the last moment, on the way out, and never appears in model context or tool arguments. Connecting an account is a short-lived link that finishes in the browser, locked to one teammate and one connector. Sending, deleting, publishing, and bulk changes wait for a person unless the teammate's instructions say otherwise.

Deployment

Hermes installs with one shell command on Linux, macOS, WSL2, or Android, and runs on a $5 VPS, a GPU cluster, Docker, SSH, Daytona, or Modal. Someone keeps that process up, applies updates, and holds the config.

Chickpea is a Deploy to Cloudflare button that clones the repository into your account, then connect Slack, pick a provider, send a message. On Cloudflare it runs on Workers with Durable Objects and D1, so nothing needs to stay alive. There is also a Node path for a host you run. On Cloudflare you can start with no API key at all through Workers AI.

Fair is fair

Where Hermes is better.

It learns. Autonomous skill creation and self-improvement during use are the headline feature, and Chickpea does not do that. Chickpea skills are imported from a URL and pinned to a commit.
Twenty platforms. Telegram, Discord, WhatsApp, Signal, Matrix, Teams, email, SMS, and more from one gateway, and a conversation can move between them. Chickpea is Slack only.
Terminal backends. Local, Docker, SSH, Daytona, Singularity, and Modal for running real work. Chickpea has connectors, MCP, repositories, and an optional Cloudflare coding sandbox.
Voice and Nous models. Incoming audio is transcribed, replies can be spoken, and Nous Portal is a first-class provider.

Questions people ask

Is Hermes Agent open source?

Yes. Hermes Agent is MIT licensed and built by Nous Research, released in February 2026. Chickpea is Apache 2.0.

How good is Hermes's Slack integration?

Better than most open source agents. It streams through Slack's native API, shows task cards, uses Block Kit, supports slash commands, and can generate the Slack app manifest. Chickpea adds native sortable tables, a live status line, user-group handles per teammate, and management by asking in Slack.

Can Hermes run several named bots in Slack?

Yes. Bot Mode creates specialist bots with their own model, memory, and skills, and one gateway can hold several bot tokens. Chickpea creates a teammate from one message in Slack and gives it its own handle, with nothing extra to configure on the Slack side.

Who can talk to a Hermes bot in Slack?

Whoever is in the allowlist of Slack member IDs in its configuration; without that setting it denies everyone. Chickpea provisions every full member of the workspace automatically the first time they talk to a teammate, and a teammate only answers in the channels it is published to.

Does Hermes ask before it acts?

Not as a built-in step. It gates who can reach it, and you can write rules into its instructions. Chickpea waits for a person before sending, deleting, publishing, or bulk changes, and any change that widens a teammate's capabilities comes back as a proposal to approve.

Can Chickpea learn new skills on its own like Hermes?

No. Chickpea imports skills from a public GitHub repository or a skills.sh link and pins the commit it inspected. Skills that ship scripts are rejected, because inline skills cannot execute imported files. That is a deliberate trade of adaptability for predictability.

Which costs less to run?

About the same: a small host plus your model provider. Hermes can run on a $5 VPS. Chickpea needs Cloudflare Workers Paid at $5 a month, or a Node host, and can use Workers AI with no provider key at all.

Try it

Your first teammate, this afternoon.

One deploy into your own Cloudflare account, connect Slack, pick a model, and ask for a teammate in a sentence. Apache 2.0, no seats, no credits.

Deploy to Cloudflare Star on GitHub

Or read how it works first.

More comparisons

Sources

Facts about Hermes Agent come from Nous Research's own pages and the coverage linked above, and were last checked on the date in the byline. Products change; if something here is out of date, open an issue and it will be fixed. A plain-text copy of this page is at /compare/hermes.md. Hermes Agent and its logo belong to Nous Research, which is not affiliated with Chickpea.