Architecture
How the desktop runtime turns model calls into reliable agent sessions: streaming loops, tool routing, crash-safe state, provider abstraction, and context management.
Home Blog
Engineering & Research
The Orkas Blog explains how Orkas works as an open-source, local-first desktop client for coordinating AI agents with your own model keys. These engineering notes cover the Agent Harness runtime, multi-agent orchestration, local self-evolution, provider routing, memory, context compaction, and the security choices behind direct-to-provider model traffic. Each article is written as a practical reference for developers, operators, and AI builders evaluating local-first AI agent workflows.
What the Orkas Blog covers
Multi-agent workflows
How a lead agent decomposes a request, dispatches sub-agents, passes context between steps, and recovers when a task fails.
Local-first AI and security
Why Orkas keeps workspace data and model keys on the user’s machine, and why traffic from your own provider goes directly to it rather than through Orkas servers.
Latest engineering articles
Product
Product
In-House, Agency, or On Your Own: How to Actually Decide
An agency sells three separable things — capacity, expertise, accountability — bundled into one invoice, and most businesses need fewer than three. Capacity has become far cheaper for one person to produce; the other two have barely moved. How to work out which one you are actually buying.
Aug 24, 2026
Product
Product
What to Charge as a Freelance Marketer (With a Free Pricing Calculator)
Start from what you need to earn, not from what other people charge. A pricing calculator that runs in your browser, the arithmetic behind it, where the real market figures actually live, and four ways to raise a rate with a client you already have.
Aug 24, 2026
Agent
Agent
Monthly Client Report Template: What to Include and How to Write It
The six sections a client actually reads, what belongs in each, and the order to actually write them in — plus a Word and Excel template already filled in, and the two-step recipe that generates both. No email required.
Aug 24, 2026
Research
Research
Bigger Was Not the Point: How Kimi K3 Scales Information Flow in Three Directions
Kimi K3 went from 1T to 2.8T parameters, and that is the least interesting number in the report. The architecture scales along sequence, depth, and width — and one change replaces an entire GPU code path with a single lower bound. What each direction buys, what it costs in the serving stack, and one industry habit the paper quietly drops.
Aug 4, 2026
Architecture
Architecture
Context Compaction Cuts by Token Count, Not by What Is Safe to Forget
Almost every agent compacts at a percentage of the window and drops the oldest turns. That threshold knows you are out of room; it knows nothing about what is safe to lose. Why the milestone Markov property is an assumption rather than a fact, and why compaction needs it to hold far more strictly than training does.
Jul 31, 2026
Architecture
Architecture
Declared Done Is Not Verified Done: Milestone Design for Long-Horizon Agents
Orkas records durable plan milestones, and separately records host-side facts about what each tool call actually changed. Nothing connects the two, so a step counts as complete the moment the model says so. How BEACON positions its detector, the three layers to build, and one criterion the paper does not have.
Jul 30, 2026
Architecture
Architecture
Loop Detection Is Not Stall Detection: Catching Agents That Spin Without Repeating
Orkas has three loop guards, and none of them catch a capable model that stalls — because all three ask whether it is repeating itself, and a stuck model never does. The blind spot, the output-side definition of progress we borrowed from BEACON, and what to measure first.
Jul 28, 2026
Research
Research
BEACON: Milestone-Guided Long-Horizon Agents
A close read of BEACON from Zhejiang University: why long-horizon agents collapse under RL, how milestone-anchored credit assignment fixes it, and a metric that seems to contradict its own design until you trace the math.
Jul 27, 2026
Research
Research
How to Get Cited by ChatGPT: What Actually Decides Whether You Get Quoted
Getting cited is not ranking. It's surviving retrieval, then being the passage worth quoting — the three OpenAI bots, the CDN gate robots.txt hides, and why a fact behind JavaScript doesn't exist.
Jul 14, 2026
Agent
Agent
How to Run Claude Code and Codex Together — One Chat to Orchestrate Both
Claude Code and Codex each win at different things. Run them together — with terminals and git worktrees, or from one Orkas Commander that orchestrates both in a single chat.
Jul 11, 2026
Architecture
Architecture
Cloud Sync in Practice: How Orkas Syncs Data Across Devices
How Orkas syncs user data across devices with encrypted transfer, content storage, server-owned commits, account locks, sync rules, model-assisted conflict handling, delete confirmation, and a recycle bin.
Jul 1, 2026
Architecture
Architecture
Rewriting the Agent's Foundation: A Ground-Up Refactor of Orkas
How Orkas rebuilt its agent foundation across the 1.0 release line: an in-process runtime, provider rotation, dynamic group-chat orchestration, open hosting, memory, and self-evolution.
Jun 25, 2026
Architecture
Architecture
Multi-Agent Orchestration in Practice: How Orkas Runs a Lead Agent and Its Sub-Agents
Inside Orkas's multi-agent orchestration: a lead agent turns one request into a plan, dispatches sub-agents by dependency, passes context between steps, and heals from failure.
Jun 16, 2026
Product
Product
Why Your AI Video Bill Is Mostly Retries
Iteration rate, not per-second price, is the dominant cost driver. Where the retries come from, and how to make the field checklist the default.
Product
Product
What Is Local-First AI? Your Data, Your Keys, Your Machine
What local-first AI means — your data, API keys, and model traffic stay on your own machine, not a vendor's cloud — why it matters for privacy, and how a bring-your-own-key agent actually works.
Jun 16, 2026
Architecture
Architecture
The Layer That Turns a Model Into a Product: Engineering Orkas's Agent Harness
How Orkas turns model calls into a reliable desktop agent runtime: streaming run loops, tool routing, context compaction, provider abstraction, memory, and crash-safe sessions.
Jun 10, 2026
Agent
Agent
An Agent That Gets Better on Its Own: Inside Orkas's Self-Evolution
Inside Orkas's local self-evolution loop: lightweight signals, background reflection, executable skills, skill metrics, and guardrails against learning the wrong lesson.
Jun 10, 2026
Orkas, explained
What is Orkas?
Orkas is an open-source, local-first desktop AI client for macOS and Windows. Instead of chatting with a single assistant, you direct a team of agents: a lead agent owns your goal and recruits sub-agents that call skills to do the work. You can use optional Orkas-managed official models or connect your own provider through OAuth or an API key; your workspace is local-first by default.
Why is Orkas local-first?
Local-first means your data and control stay on your device. Your chats, files, knowledge base, memory, and encrypted model keys live on your machine by default. With your own provider, model traffic goes directly from your computer to that provider and is not proxied through Orkas servers; official models use Orkas's managed model service.
How do teams of agents work in Orkas?
A lead agent reads your goal, breaks it into steps, and recruits sub-agents by task and capability. Each sub-agent works in its own bounded context and calls skills — web search, code execution, file I/O, knowledge-base search, and connectors — to deliver. The lead passes each sub-agent only what it needs, which keeps token cost down and responsibilities clean.
What can you use Orkas for?
Common uses include research and analysis, writing and editing, coding with native or external CLI agents such as Claude Code, Codex, and OpenClaw, data work, learning, and office documents. You can also turn a recurring task into a reusable sub-agent once and summon it in chat whenever you need it.
What are Orkas's limitations?
The current public release is a macOS and Windows desktop app; the iOS remote-control relay is disabled, and there is no web client. You can use optional Orkas-managed official models or connect a provider through OAuth or an API key. Your provider bills usage from your own key; managed capabilities use Orkas credits. Output quality depends on the model you connect. Optional multi-device sync stores synced data on Orkas servers, and the free edition sends limited usage analytics.