Skip to main content
AIDiveForge AIDiveForge

Hermes Agent vs Strands Shell

Which one should you try?

Hermes Agent

Free

Best for People who want an assistant that remembers context across sessions

Visit Hermes Agent See full review

Strands Shell

Free

Best for Python and TypeScript developers building agentic workflows

Visit Strands Shell See full review

Field notes

All field notes →

Pricing, capabilities and platforms side by side. Prices are as of the "Pricing checked" date in the table.

Hermes Agent

Hermes Agent

Hermes Agent from Nous Research keeps memory across sessions, saves reusable skills and can be reached through messaging apps. We ran it on three tasks with a local 3B model and with Claude Opus 5.5.

Strands Shell

Strands Shell

The core pattern is tight: decorate a Python or TypeScript function with `@tool`, pass it to an `Agent`, attach hooks that fire before or after each tool call, and the agent runs its loop. The `BeforeToolCallEvent` hook lets you inspect the tool's name and input — and cancel the call with a message if your conditions aren't met. That's not a workaround; it's the documented pattern. Where the framework gets quiet is multi-agent coordination — the docs describe single-agent tool loops clearly, but teams building agents that hand off to other agents will find precious little guidance on failure recovery between hops. When that gap bites, teams layer their own orchestration logic on top, which means maintaining that logic themselves.

AttributeHermes AgentStrands Shell
PricingFreeFree
PriceFree (MIT, self-hosted)—
Pricing checkedJul 1, 2026Not checked yet
Free trialNot statedNot stated
Open sourceYesYes
Has APIYesYes
Self-hosted optionYesYes
PlatformsmacOS, Linux, Windows (native or WSL2), Android (Termux); runs commands locally or in Docker, SSH, Singularity, Modal, Daytona, Vercel SandboxPython, TypeScript, cross-platform
Released2026-022025-05
Pros
  • Memory and saved skills carry over between sessions.
  • Reach it from a CLI or messaging apps such as Telegram and Discord.
  • You choose where commands run, including Docker, SSH and hosted sandboxes.
  • Passed all 9 of our tasks with Claude Opus 5.5 and contacted only Anthropic's API while doing it.
  • Pre-tool hooks (`BeforeToolCallEvent`) let you inspect and cancel any tool call before it executes, so enforcement rules — citation requirements, output validation, safety checks — live in one function rather than scattered across prompt engineering.
  • Model-agnostic design means switching the underlying LLM provider is a config-level change, so you are not rewriting tool definitions or hook logic when API costs shift or a new model performs better on your task.
  • Apache-2.0 license with a self-hosted path means no managed-service dependency and no vendor lock-in on the runtime — teams with data-residency requirements can run the full stack on their own infrastructure.
  • Typed tool definitions (Python type annotations, Zod schemas in TypeScript) give the agent a machine-readable input contract, which reduces malformed tool calls and makes testing individual tools straightforward without spinning up a full agent.
  • Separate evaluation tooling (`strands-agents/evals`) ships alongside the core SDK, so teams can measure agent behavior against defined criteria rather than eyeballing outputs — which is the difference between shipping with confidence and shipping with hope.
Cons
  • Passed 0 of 9 tasks with a 3B local model; it needs a large model and a 64K context.
  • Heavy on tokens: an ~11,500-token prompt on every call and about 47,100 tokens per task, six times Goose's.
  • It installs its Anthropic support from PyPI on first use, so offline setups fail until you pre-install it.
  • Multi-agent handoffs — where one agent's output becomes another agent's input and something fails mid-chain — are not addressed in the documented patterns. Teams building that architecture write their own recovery logic on top of the SDK, and at the point where that logic grows, they are maintaining a custom orchestration layer that the framework does not help them test or observe.
  • The hooks model fires at tool-call boundaries, which covers pre- and post-tool enforcement cleanly. It does not expose mid-reasoning interception — if you need to inspect or redirect the model's chain-of-thought before it selects a tool, there is no documented hook for that. Teams that need reasoning-level control end up wrapping the model call themselves.
  • Teams that hit the limits of single-agent tool loops at scale — specifically those needing stateful, multi-agent pipelines with built-in retry semantics and distributed execution — report moving to frameworks like LangGraph or Temporal-backed orchestration, where those primitives are built in rather than delegated to the team.

Made up your mind?

Hermes Agent

Free

Best for People who want an assistant that remembers context across sessions

Visit Hermes Agent See full review

Strands Shell

Free

Best for Python and TypeScript developers building agentic workflows

Visit Strands Shell See full review

Related comparisons

Free trials & curated lists

Specs come from each tool's own site and docs (how we test). AIDiveForge is editorially independent.

Share this comparison

Local AI tools we tested

Browse tools · Affiliate Disclosure · Privacy · Cookies · · Email Policy · Terms · DMCA