Compare·Self-improving autonomous agent
Continuum
Multi-agent workbench
VS
Hermes logo
Hermes Agent
Self-improving autonomous agent

Continuum vs Hermes Agent

An agent that remembers you versus an instrument for shipping reviewed code. Both are real products; the jobs differ.

Hermes Agent is Nous Research’s MIT-licensed self-improving agent, released 25 February 2026 and past 225,000 GitHub stars by August 2026. It runs as a persistent daemon on infrastructure you own (a $5 VPS, a GPU cluster, or serverless on Daytona or Modal), keeps memory across sessions, writes and refines its own skills, schedules cron work in plain language, spawns isolated subagents, and reaches you on 20+ channels including Telegram, Discord, Slack, WhatsApp, Signal, email, and the CLI. Continuum is a narrower instrument on purpose: a supervised coding workbench that hosts the official Claude Code and Codex CLIs under your own subscriptions, with git worktrees, a diff and PR review path, live 5h and weekly quota gauges, multi-account isolation, and a local dollar ledger by repo. The honest split is persistence versus supervision, not smart versus dumb. Background reading: AI coding agents.

Download for Mac All comparisons
Workflow Monday path Matrix Decide Money FAQ

Updated 2026-08-03 · Mac stable · Win/Linux desktop beta

Choose Continuum when

Your week is repos, branches, and pull requests: you want official Claude and Codex under your own plans, worktree isolation, a diff you actually read, quota gauges, and spend truth per repo.

Choose Hermes Agent when

You want a personal agent that stays running, remembers your projects between sessions, messages you on Telegram or Slack, and improves its own skills over time.

Snapshot adjacent job
Dimension Continuum Hermes Agent
Job-to-be-done Supervised multi-provider coding Always-on personal agent that learns
Runs as Desktop app + daemon on your machines Persistent daemon on VPS, cluster, or serverless
Memory Per-session transcripts, no cross-session learning Persistent memory + self-written skills
Reach Native iPhone + Watch · web · QR/Tailscale pair 20+ chat channels incl. Telegram, Slack, WhatsApp
Coding depth Worktrees · diff staging · PR pane · official CLIs Can code; no git worktree or PR product
Autonomy Human-gated plan and diff by default Scheduled, autonomous, self-improving loop
Cost view Local JSONL → $ by repo/provider/day Model credits + whatever keeps the daemon up
Mac stable · web · iPhone · Watch · Win/Linux desktop beta · free app
01

Always-on agent vs supervised coding ops

In Continuum

Ship day

09:00

Claude in a worktree on a customer bug, Codex in another on regression tests. The sidebar is the ops board and both are on your own plans.

11:30

Plan-ready hits the phone. You approve from the queue at standup; the Mac keeps the session live and the diff waits for you.

14:00

Diff pane, stage the three hunks you believe, discard the rest, open the PR from the same tab. Nothing merges that you did not read.

17:30

Ledger by repo shows the week’s cost. Hermes, if you run it, stays on its own box with its own credentials and its own budget.

In Hermes Agent

Always-on agent day

07:00

Hermes has already run the scheduled morning briefing and dropped it in Telegram before you opened a laptop.

11:00

You message it from Slack about a recurring report. It remembers last month’s format because memory persists across sessions.

15:00

It writes a reusable skill out of the run, refines it on the next invocation, and the improvement sticks. Continuum has nothing equivalent.

19:00

The daemon hibernates on serverless and costs close to nothing idle. Repo work with a review gate is still a different tool’s job.

02

Monday path

Monday for shipping code, with Hermes in the other lane

Install, connect, first session - steps you can run the same day.

01

Install Continuum on the Mac and attach the Claude and Codex accounts you already pay for.

02

Spawn worktree sessions per ticket so two agents never fight over the same working tree. See worktrees for Claude Code.

03

Keep plan mode on for anything touching production branches; Code mode is for work you will review in the diff pane anyway.

04

Pair the iPhone so mid-day plan-ready lands on your phone instead of a chat channel you will scroll past.

05

Check the weekly Claude rail before kicking off a night batch, and pin the secondary account if the primary is hot.

06

If Hermes runs beside it, give it its own VPS or container and its own least-privilege credentials, never the repo tokens Continuum uses.

07

Let Hermes own the scheduled briefing; let Continuum own anything that ends in a pull request.

08

Friday: read cost by repo and decide what next week’s experiment budget actually is.

How Continuum runs production coding agents

How Continuum runs production coding agents

  • Official CLIs under your own logins: Claude over PTY, Codex over the ACP harness, plus Cursor, Grok, Gemini, and OpenCode. No self-modifying runtime in the default path.
  • Worktree isolation, a plan gate, per-hunk diff staging, and a PR pane decide what actually reaches main.
  • A pre-prompt git checkpoint is taken before each turn, so an agent’s bad idea is recoverable rather than archaeological.
  • Mobile pairing runs over Tailscale or loopback with a bearer token; the daemon refuses peers outside those ranges.
  • Sends are rate-limited and mobile commands are idempotent, so a flaky phone network cannot double-submit a prompt.
  • Audit logs record sends, config swaps, and mobile commands as hashed JSONL for accountability after the fact.
  • Local analytics price session history against a pricing snapshot and answer cost questions without a telemetry backend.
03

Adjacent agent matrix

Cells use product nouns on both sides. Hermes owns persistence, learning, and reach; Continuum owns repo-shaped review, quota, and cost. A row marked no is a scope decision, not a gap someone forgot to close.

Capability Continuum Hermes Agent
Official Claude Code / Codex hosts First-class, under your subscription Model-agnostic via Portal, OpenRouter, OpenAI
Persistent cross-session memory Transcripts only, no learned memory Core feature, searchable + summarized
Agent writes its own skills Reads skills you author Closed learning loop + Skills Hub
Chat-app reachabilityTelegram · Slack · WhatsApp · Signal Own apps and web, not chat bridges 20+ channels through one gateway
Scheduled / cron work Scheduled session follow-ups Natural-language schedules
Git worktree isolation Worktree per session + spawn grid Not a worktree product
Diff review + PR pane Per-hunk staging, PR status, merge No PR review surface
Pre-turn safety checkpoint Git checkpoint before each prompt Sandbox backends: Docker, SSH, Modal
Live 5h / weekly quota gauges Menu-bar rails per provider Credit balance, not plan windows
Multi-account subscription isolation CLAUDE_CONFIG_DIR / CODEX_HOME isolation One key set per instance
Local repo $ ledger Local JSONL → $ by repo/provider/day Portal credits and host bill
MCP servers Through the hosted CLIs; inventory in Settings Native MCP client
Free and open source Free app, commercial product MIT licensed
04

What each refuses to be

01 · Persistence

It remembers. Continuum does not.

Hermes keeps memory across sessions and turns successful runs into skills it reuses and improves. Continuum keeps transcripts and analytics, and deliberately learns nothing about you between sessions. If you want an agent that compounds, that is Hermes’ lane and it is a real advantage.

02 · Supervision

Continuum optimizes for the review gate.

Plan mode, per-hunk diff staging, a pre-prompt git checkpoint, and a PR pane exist so a human decides what lands. Beside a self-improving agent this looks conservative. It is meant to. See reviewing AI-generated code.

03 · Reach

Chat channels versus native clients.

Hermes reaches you wherever you already message: Telegram, Slack, WhatsApp, Signal, email, SMS. Continuum reaches you through its own iPhone, Watch, and web clients, which is narrower but carries plan, diff, PR, and terminal panes a chat bubble cannot.

04 · Rails and money

Plan windows are Continuum’s obsession.

Hermes bills credits. Continuum instruments the 5h and weekly windows of the Claude and ChatGPT plans you already own, stacks multiple accounts with config-dir isolation, and prices every session locally by repo, provider, and day.

05 · When Hermes wins

If the agent should outlive the session.

Ambient automations, scheduled briefings, an assistant that learns your projects and answers from your pocket app of choice: pick Hermes. Continuum will feel like an IDE, correctly, because that is what it is.

05

Coexistence

Stack recipe

How people run both.

Stack recipe: run Hermes for the always-on lane (briefings, scheduled reports, ambient automations, chat-app reachability) and Continuum for the shipping lane on Claude and Codex. They can share a machine. Keep Hermes off production main until you trust its loop, and keep production secrets out of any agent that rewrites its own skills. Continuum makes no claim to Hermes’ memory or self-improvement.

Download Continuum Providers
06

Persistent agent or coding workbench

Continuum

Choose Continuum for reviewed, multi-provider shipping

  • Your output is pull requests, not chat threads.
  • You want official Claude and Codex under subscriptions you already pay for.
  • You need worktree isolation so parallel agents cannot collide.
  • You want a diff you stage by hunk before anything reaches main.
  • You run more than one Claude or Codex account and need clean isolation.
  • You want quota rails and per-repo spend without standing up telemetry.
Hermes Agent

Choose Hermes for a persistent agent that learns

  • You want an agent that keeps running when you close the laptop.
  • Cross-session memory and self-written skills are the point, not a nice-to-have.
  • You want to message the agent from Telegram, Slack, WhatsApp, or email.
  • Scheduled briefings and ambient automations matter more than PR review.
  • You are happy on a self-hosted daemon with your own credits and sandboxing.
  • You prefer an MIT-licensed stack you can fork.
07

What you pay

Continuum

App + your labs

$0 for the app on Mac, web, iPhone, and Watch. Bring your own keys or the Claude and ChatGPT plans you already pay for, and Continuum takes no cut. Optional hosted inference: Plus $25/mo, Max 100 $100/mo, Max 200 $200/mo, Ultra $500/mo.

Hermes Agent

Their bill

Hermes Agent itself is MIT licensed and free. You pay for inference and for keeping the daemon alive: self-host on a $5 VPS or serverless (Daytona, Modal) where it hibernates near zero when idle, or use Hermes Cloud, launched July 2026 and billed hourly from credits. Models come from Nous Portal, OpenRouter, OpenAI, or a custom endpoint. Nous Portal has a free tier plus paid tiers (Plus, Super, Ultra) that bundle monthly credits and 300+ models; Plus is listed at $20/mo as of August 2026.

How to compare

Total cost of work

Compare the whole lane, not the license. Hermes is free to install and metered on credits plus hosting; Continuum is free to install and rides plans you already own, with hosted inference only if you want it. See AI coding pricing compared.

Hermes Agent plans checked 3 August 2026 against their pricing page. Vendors reprice often - if you spot a stale figure, tell us and we will fix it. Continuum never marks up a session it runs under your own login.

Continuum ladder: Free app · Plus $25/mo · Max 100 · Max 200 · Ultra - full pricing. Also analytics, multi-account, devices.

08

Questions

Continuum vs Hermes Agent.

Deep dives: docs, providers, sessions.

An MIT-licensed self-improving agent from Nous Research, released 25 February 2026. It runs as a persistent daemon on your own infrastructure, keeps memory across sessions, writes and refines its own skills, schedules work in natural language, spawns isolated subagents, connects to MCP servers, and can be reached from 20+ chat channels.

The agent is free and MIT licensed. Your bill is inference plus hosting: a $5 VPS, serverless that hibernates when idle, or Hermes Cloud billed hourly from credits. Nous Portal has a free tier and paid tiers that bundle credits and 300+ models.

Yes, and it is genuinely capable at it. What it does not ship is the repo-shaped review layer: git worktrees per task, per-hunk diff staging, a PR pane with status and merge. That layer is what Continuum is for.

Only if you misread the category. For supervised multi-provider coding with review gates and cost truth, Continuum is the tool. For a persistent agent that learns and reaches you anywhere, Hermes is a different product and this page is not trying to talk you out of it.

No. Continuum keeps session transcripts and local usage analytics and deliberately builds no cross-session model of you. If an agent that compounds over months is what you want, that is Hermes’ design center, not Continuum’s.

Yes, and most people who use both keep them apart on purpose: Hermes on its own host or container with least-privilege credentials, Continuum on the workstation holding your repo tokens. Different lanes, different blast radius.

Continuum, by construction. Plan mode, a git checkpoint before each turn, per-hunk staging, and a human on the PR pane are the default path. An autonomous loop on production main needs gates you build yourself. See AI coding agent security.

Both are MIT-licensed, self-hosted, always-on personal agents you message from chat apps. Hermes leans on a closed learning loop, self-written skills, and Nous Portal models; OpenClaw leans on its gateway, channel breadth, and native companion apps. Continuum sits beside either as the coding workbench, and there is a Continuum vs OpenClaw page too.

No. Continuum’s provider set is mainstream coding CLIs: Claude, Codex, Cursor, Grok, OpenCode, and Gemini or Antigravity. Hermes is compared here as an adjacent job, not claimed as a Continuum engine.

Further reading

Guide

How to run multiple Claude Code sessions in parallel

Two to six agents at once: the isolation, the rate-limit math, and how to review the output without drowning.

Read guide → Guide

Claude Code plan mode: how to use it and when it pays off

Read-only research first, edits only after you approve. The single best habit for work that matters.

Read guide → Guide

Git worktrees: the complete guide

One repository, many working directories. The mechanics, the full command surface, and the six things that go wrong.

Read guide →
Begin

Run your agents
in Continuum.

Free app. Your subscriptions. Optional hosted inference. Mac stable - Windows and Linux desktop are beta.

Download for Mac See pricing

vendor-neutral · local-first · multi-device