CARAKA vs OPENCLAW vs HERMES

They are personal assistants that can also write code. Caraka only writes code.

This is not a fight over the same users. OpenClaw clears your inbox, keeps your calendar, and drives your browser. Writing code sits somewhere on that list. For Caraka it is the entire list.

“

OpenClaw is a personal agent; OpenCode is a coding agent. Different tools for different jobs.

Standard Compute, 2026

OpenClaw

PERSONAL AI ASSISTANT

A persistent, self-hosted assistant that connects a model to your OS, your messaging apps, and the internet. It acts rather than advises.

·Clears your inbox·Manages your calendar·Drives a browser·Runs shell commands·Writes code, among other things

Hermes

AGENT FRAMEWORK

A Python agent framework with its own runtime and orchestration. You build the assistant you want on top of it.

·Orchestrates agents·Ships its own tool layer·Linux, macOS, WSL2 only·Writes code, among other things

Caraka

CHAT BRIDGE TO YOUR CODING AGENT

A bridge with no runtime of its own. It carries your task to the coding agent you already installed and already pay for.

·Works on your repository·Sessions as topics, threads, or a header·Approval before every write·An append-only audit log·Nothing else, on purpose

01 — SIDE BY SIDE

Where the two shapes actually differ.

OPENCLAWHERMESCARAKA

CategoryPersonal assistantAgent frameworkChat bridge

Main jobInbox, calendar, browserOrchestrationYour repository

Agent runtimeIts ownIts ownYours

Execution tools20+YesZero

Plugin marketplaceLargeYesNone

Channels22via gateway3 in v0.6

Tokens outside your agentYes, own loopYesNone

SandboxingConfigure itConfigure itInherited

Reported setup timeHourspip + venvTarget: under 3 min

MaturityVery highHighv0.6 closed beta

Skywork measured usage distribution across both products and found 85% of OpenCode's usage is software development, while OpenClaw's spreads across personal use. The two products serve different people.

02 — WHAT USERS REPORTED

Six complaints that shaped this project.

These come from a reviewer who used OpenClaw daily from January 2026 for both personal and professional work, and liked it enough to write at length before switching. Some may already be fixed. The dates are printed so you can judge.

SETUP

“It is open-source, yes, but installation is a nightmare. I remember spending hours to get it right.”

Composio, May 2026

→npx caraka init. One command, four interactions, target under three minutes. The wizard confirms what it finds rather than asking what it should already know.

PLUGINS

“The Skills and Plugin ecosystem is vast, but the experience is horrible.”

Composio, May 2026

→No marketplace at all. Nothing loads dynamically, and runtime dependencies are pinned in package-lock.json. Trust sits with you, not with a registry.

COST

“Agentic tasks consume a massive amount of tokens. If you want to use it like a personal assistant, the cost will skyrocket pretty fast.”

Composio, May 2026

→Caraka has no agent loop, so it burns no tokens of its own. Everything runs on the coding-agent subscription you already pay for.

SECURITY

“The initial releases lacked any security hardening. The internet was filled with heated debates around it.”

Composio, May 2026

→Mandatory allowlist, signed single-use approvals, secret scrubbing, and an append-only audit log, all present from the first commit rather than added after an incident.

STABILITY

“Every patch would make it worse. A fix for one thing would unfix the other.”

Composio, May 2026

→A core capped at roughly 8,000 lines is small enough that one person can read it in a day. Fewer moving parts is the only reliable way to make a patch predictable.

ARCHITECTURE

“Not bugs, since the project ships fast and breaks fast in equal measure. The edges I hit were architectural.”

Vellum, August 2026

→The written non-goals are the defence. No agent loop, no tools, no marketplace, no companion app. Every feature request is tested against one question: can the coding agent already do this?

03 — THE TOKEN MATH

An assistant with its own loop bills you twice.

This is the clearest economic difference, and it follows directly from architecture rather than pricing policy.

ASSISTANT WITH ITS OWN LOOP

your coding agent plan$20+/mo
assistant reasoning looptokens
heartbeat every 30 mintokens

Every proactive turn is a full agent turn. The bill arrives from two directions.

CARAKA

your coding agent plan$20+/mo
caraka bridgefree
heartbeatoff

No loop, no second bill. Proactive scheduling is off by default because it is expensive and noisy.

04 — WHICH ONE

Pick by the job, not by the feature count.

Clear my inbox, keep my calendar, drive a browserOpenClaw

Build my own assistant on a Python frameworkHermes

Ship a fix to my repository from my phoneCaraka

Keep five tasks running without them colliding in one chatCaraka

Approve every write before it landsCaraka

Not pay for a second reasoning loopCaraka

A mature project with a large community todayOpenClaw

WHAT WE WILL NOT CLAIM

·OpenClaw is far more mature. Hundreds of thousands of stars, tens of thousands of commits, a real community. Caraka v0.6 is an early preview and you should treat it that way.

·OpenClaw does work Caraka will never do. Calendar, inbox, browser, home automation. If that is what you need, this is not the answer.

·The complaints quoted above come from 2026 releases and some may already be fixed. Dates are printed so you can check for yourself rather than take our word for it.

Send the task. Caraka runs it.

v0.2 carries the task to Claude Code. The free Gemini CLI path is roadmap work. Caraka itself never asks for a card.

Read the docs →