Hermes Agent vs Claude Code vs OpenCode: which coding agent should run your terminal in late 2026?
Three agents, three philosophies: a self-improving generalist with a messaging gateway, a vendor-tied CLI shipping daily, and an MIT-licensed coding TUI. An Axion Scorecard on what we can actually verify.
Hermes Agent canary 20261020T065957Z. The self-improving AI agent built by Nous Research.
Comprehensive take
Stay on Claude Code if you are already in the Anthropic ecosystem; pick Hermes for a model-agnostic, memory-bearing agent across chat apps; pick OpenCode for an MIT coding TUI with a safe read-only plan mode.
Use cases
Personal ops agent across chat apps
Hermes Agent's single gateway to Telegram, Discord, Slack, WhatsApp and Signal plus a cron scheduler fits always-on personal automation; no equivalent in the other two.
Enterprise permission governance
Claude Code v2.1.304's `decider: "person"` and `e.isReadOnly` give platform teams hooks to force human approval on risky tool calls.
Audit-first code exploration
OpenCode's `plan` agent denies file edits by default and asks before bash, a sane starting mode for unfamiliar repos.
Specs
| Hermes Agent licence | MIT source |
|---|---|
| Hermes Agent terminal backends | 7 (local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox) source |
| Hermes Agent GitHub stars | 253.6k source |
| Claude Code version / date | v2.1.304, 19 Oct 2026 source |
| Claude Code skill hook cap per source | 1,024 hooks / 16,384 matcher chars source |
| OpenCode licence | MIT source |
| OpenCode GitHub stars | 214.6k source |
Insights
- Hermes is the only one of the three that is both model-agnostic and MIT licensed with cross-session memory.
- Claude Code's v2.1.304 fix list (worktree deletion, compaction window parsing) shows the product under heavy real-world load.
- No vendor publishes like-for-like agent benchmarks on these release pages; Capability is unscored.
Axion Scorecard
Overall: 7.0
Not yet scored — no cited data
Hermes and OpenCode are MIT; Claude Code is proprietary and single-vendor. source
OpenCode has the broadest installer set; Claude Code the richest hook/permission API; Hermes the most surfaces. source
Hermes and OpenCode are free software; Hermes bundled services need a Nous Portal subscription or own keys; Claude Code needs an Anthropic plan. source
Claude Code adds human-only deciders and read-only flags; OpenCode plan mode denies edits; Hermes has 7 sandbox backends. source
Hermes' self-improving skills and user modelling are the most novel surface; Claude Code's Code Review levels are incremental. source
Why compare these three
All three shipped in the last 96 hours. Nous Research's Hermes Agent cut its fourth daily canary in a row (v0.22.2+canary.20261020T065957Z, commit c5a3c0f, tagged 20 October). Anthropic's Claude Code released v2.1.304 on 19 October with over 80 changelog entries. OpenCode, from anomalyco, tagged v2.1.0 on 20 October. For a practitioner choosing where to spend the next quarter's muscle memory, these are the live options.
A note on method: the Hermes canary release page rendered no notes at fetch time (only the tag message), and the OpenCode release page failed to load. Where we cite Hermes and OpenCode, the facts come from their repository READMEs. Where we cite Claude Code, they come from the v2.1.304 release notes. We score only what we could read.
Hermes Agent: the generalist with memory
Hermes describes itself as 'The self-improving AI agent built by Nous Research.' It is MIT licensed, installs with a one-line shell script, and is model-agnostic: Nous Portal, OpenRouter, OpenAI, custom endpoints, switchable with `hermes model`. The README claims 300+ models via Nous Portal.
What distinguishes it is not coding specifically. It ships agent-curated memory, FTS5 session search with LLM summarisation, Honcho-based user modelling, autonomous skill creation compatible with the agentskills.io standard, 40+ tools, isolated subagents, MCP integration, a cron scheduler, and a single gateway to Telegram, Discord, Slack, WhatsApp and Signal. Sandboxing spans seven terminal backends: local, Docker, SSH, Singularity, Modal, Daytona and Vercel Sandbox. The repository shows 253.6k stars, 55.0k forks and 52,346 commits on main.
Caveats from its own README: Windows antivirus may quarantine `uv.exe` as a false positive; web search, image generation, TTS and cloud browser route through a Nous Portal subscription unless you bring your own keys.
Claude Code: vendor-bound, shipping daily
Claude Code v2.1.304 is a dense release. Practitioner-relevant additions include `decider: "person"` in a mod's `tool.check` answer (so only a human or host callback can approve certain tool asks), `e.isReadOnly` on tool-check events, `/auto-mode-setup` now on Bedrock, Vertex and Foundry, and Code Review levels (Quick, Standard, Deep) per organisation or repository.
The fixes list is also telling about where the product is being stressed: `claude --worktree` could remove a worktree and branch without asking on detached HEAD, 'which could lose commits'; `CLAUDE_CODE_AUTO_COMPACT_WINDOW=200k` silently set a 100K window; skill hooks are now capped at 1,024 hooks and 16,384 matcher characters per source. These are the bugs of a tool used hard at scale. Claude Code is not open source and runs Anthropic models only, which is the structural trade.
OpenCode: the MIT coding TUI
OpenCode calls itself 'The open source AI coding agent.' MIT licensed, installable via `curl`, npm (`@opencode/cli`), Homebrew, Scoop, Chocolatey, pacman and Nix, with a beta desktop app for macOS, Windows and Linux. It ships two built-in agents toggled with Tab: `build` (full access) and `plan` (read-only, 'Denies file edits by default', 'Asks permission before running bash commands'), plus a `general` subagent for complex searches. The repository shows 214.6k stars, 28.8k forks and 20,531 commits on dev. The README excerpt we read did not list providers, so we do not score that.
Head-to-head
| Dimension | Hermes Agent | Claude Code | OpenCode |
|---|---|---|---|
| Licence | MIT | Proprietary | MIT |
| Model choice | Any (300+ via Nous Portal) | Anthropic only | Not stated in README read |
| Memory across sessions | Yes (FTS5 + Honcho) | Compaction summaries | Not stated |
| Messaging surfaces | Telegram, Discord, Slack, WhatsApp, Signal, CLI | Terminal, VS Code, Cowork, Slack via Claude Tag | Terminal, desktop beta |
| Sandbox backends | 7 | Built-in sandbox | Plan agent restrictions |
| Release cadence (observed) | Daily canary | Near-daily | v2.0.32 (16 Oct) to v2.1.0 (20 Oct) |
Should you switch?
If you already pay for Claude and want the most polished permission model, the `decider: "person"` and read-only hook signals in v2.1.304 are genuinely ahead; stay. If you need one agent that lives in your chat apps, remembers you, and lets you swap models or run on your own endpoint, Hermes is the only one of the three built for that, and MIT means you can audit it. If you want a focused coding TUI without vendor lock-in, OpenCode's plan/build split is a clean default and the installer story is the broadest.
None of this is benchmarked. We found no public, like-for-like SWE-bench or Terminal-Bench numbers for these exact versions on their own pages, so Capability is unscored for all three. We will revisit when any vendor publishes them.