A chat
In a chat you do the work. Here the work runs itself, and anything dangerous waits for your yes. A chat gets switched off without asking; here you change the model yourself.
SANKORD among AI agent products
How SANKORD is built, next to the others
The incumbents sell a company-wide layer on their own platform; SANKORD puts the agents inside one business or one department, within its walls – a difference of level, not of price.
This page sets SANKORD beside twelve other products that run AI agents: where each one runs, who it is for and what it gives. Every fact about another system comes from its vendor's own page or a published report, named under the system with its date, so you can check every line.
Current as of October 2026
How SANKORD is built
Which tool for what
Most of the products on this page are made for people who write code, or run on the vendor's computer. SANKORD aims at the top right of the chart: made for the one who runs the business, and running on your own hardware.
A cloud agent
The SANKORD system runs inside your walls, on your servers; nothing irreversible happens without your yes; models change per task, and with open models on your own hardware no model vendor can switch you off. Their edge: starting in a minute. Our answer: you start with a trial beside your people, and nothing goes out.
The next hire
The system takes the work before you hire the next person; your people stay and decide what to hand over. Their edge: judgement. Our answer: that is why you decide.
A free tool you install yourself
Tools like these are built for people who write code. SANKORD is built for the one who runs the business.
A platform for a whole company
They sell a layer across a whole company, on their own platform. We put the agents inside one business, within its walls.
Handing the work to the cloud; hiring the next person for it.
Where each product stands and who it is for
Each product is placed by where it runs, from the vendor's servers to your own hardware, and by who it is made for, from developers to business owners. Point at or tap a name in the key to find its dot.
● Coding agents. ■ Cloud agents. ▲ Self-hosted workspaces.
SANKORD
The agents work inside your walls: on your own servers or in your own cloud account
Where the agents run and the data lives On the vendor's servers → Vendor's and your own → On your own hardware
Who it is made for For developers ↑ For people and teams ↑ For business owners
Shape shows the kind of product, colour tells products apart, and the ring marks where SANKORD aims.
- Claude Code
- Codex
- Grok Build
- DeepSeek Harness
- MiMo Code and MiMo Desktop
- OpenHands
- Osaurus
- Grok Bot
- Dots
- Octop
- OpenClaw
- Paperclip
Three layers on the market
Agent products sit in three layers. A layer is set by who works with the product and what it works on.
Coding agents
A developer's tool at the keyboard, sold together with the use of AI models. It works on files, changes, tests and pull requests. In this comparison: Claude Code, Codex, Grok Build, DeepSeek Harness, MiMo Code, OpenHands, Osaurus.
In plain words: tools for the person who writes the code.
Cloud agents
A computer run by the vendor, always on, reached through a subscription. It works on tasks, mail and a browser session, and keeps working when the laptop is closed. In this comparison: Grok Bot, Dots.
In plain words: an assistant that lives on the vendor's computer and is reached through a subscription.
Company systems
By design. The agents work inside your walls: on your own servers or in your own cloud account. Nothing irreversible happens without your yes: no payment, signature, deletion or publication. Every step is written down. One command stops the whole system. In this comparison: SANKORD.
In plain words: a system that runs on your side and is made for the one who runs the business.
Self-hosted workspaces (Octop, OpenClaw, Paperclip) run on a local machine: Octop hands work to coding agents, and Paperclip takes a coding agent.
What each system gives, can and cannot do
What each system gives the person it is aimed at, what it can and cannot do, and the dated pages each line rests on.
Claude Code
A program for developers that runs in a terminal or a desktop tab and works with Claude models.
For an owner: Aimed at a developer working on a local machine, with the vendor's own models. Its subject is files, changes and tests, as the layer description above says.
Claude Code runs in a terminal or a desktop tab and can be extended with add-ons of your own. It works with Claude models, and its default Opus is Opus 5.5 from Claude Code 2.1.280 (22 Sep 2026).
- Can
- Reads and writes files, runs commands and schedules tasks. Commands run in a protected space on macOS, Linux and, on Windows, through Windows' built-in Linux layer (WSL2).
- Cannot
- Does not work with models from other vendors. On Windows without that layer, commands run outside the protected space.
Sources: Claude Code, sandboxing (Read 6 Oct 2026); Opus 5.5 as the default Opus in Claude Code 2.1.280 (22 Sep 2026); Vendor pages as read on 2 October 2026 (2 Oct 2026)
Codex
A developer's program: a command line on the developer's computer, with tasks that can also run in parallel in OpenAI's cloud.
For an owner: Aimed at a developer. It runs on the local machine and in the vendor's cloud, with OpenAI models by default, and the code of its command-line program is open.
Codex is a program for a developer: it works in a command line on the developer's computer and can run tasks in parallel in OpenAI's cloud; the code of its command-line program is open.
- Can
- Reads and writes files, runs commands and hands tasks to the cloud. By default it has no internet access while it works.
- Cannot
- Other vendors' models are not connected out of the box: they are set up by hand in its configuration, with OpenAI as the default.
Sources: Codex CLI, OpenAI Developers (Read 9 Oct 2026); Codex Cloud, OpenAI Developers (Read 9 Oct 2026); Codex Cloud internet access, OpenAI Developers (Read 9 Oct 2026); Codex IDE extension, OpenAI Developers (Read 9 Oct 2026); Codex repository, OpenAI on GitHub (Read 9 Oct 2026); Codex advanced configuration, OpenAI Developers (Read 9 Oct 2026)
Grok Build
The grok command in a terminal: a developer's tool that accepts Claude Code's command options.
For an owner: Aimed at a developer on a local machine, with Grok models only. The client is closed.
Grok Build is the grok command in a terminal. It accepts Claude Code's command options, so a developer who already built add-ons for Claude can carry on working the same way. The client is closed, and the docs are dated July to August 2026.
- Can
- Reads and writes files, runs commands, can run unattended, runs workflows and multi-step goals, makes images and other media and brings in Claude Code sessions.
- Cannot
- Does not work with models other than Grok.
Sources: Grok Build, CLI reference (As cited in a note of 5 Oct 2026); Grok Build, modes and commands (As cited in a note of 5 Oct 2026)
Grok Bot
Named bots on a persistent cloud computer with a browser, files and a terminal, reached by subscription.
For an owner: Aimed at a subscriber. It runs on the vendor's computer, that is, on the vendor's side, and has been in beta since 11 Aug 2026.
Grok Bot gives a subscriber named bots on a persistent cloud computer with a browser, files and a terminal, plus team bots. Work continues after the laptop is closed. It has been in beta since 11 Aug 2026; secondary pages differ on the bundle, so no price is given.
- Can
- Takes voice input. A bot stops before sending, publishing, buying, entering a password or passing a captcha, so drafts can wait before mail leaves.
Sources: Grok Bot, secondary brief on vanja.io (Read 6 Oct 2026); Grok Bot, secondary brief on scalr.sg (Read 6 Oct 2026)
Dots
An always-on agent with its own cloud computer and browser, reached from ChatGPT, Slack and Teams.
For an owner: Aimed at subscribers to Pro, Business Premium and Enterprise plans. It runs on the vendor's computer and works for as long as the plan is active.
Dots is an always-on agent on GPT-6 Astra with its own cloud computer and browser, more than 4,000 app connectors, and access from ChatGPT, Slack and Teams. It launched on 29 Sep 2026 for Pro, Business Premium and Enterprise plans; teams of dots are announced, not shipped.
- Can
- Works for as long as the plan is active, takes tasks without a local install, and sets custom rules for what needs approval.
Sources: OpenAI launches Dots (TechCrunch) (29 Sep 2026); OpenAI launches Dots (MacRumors) (29 Sep 2026)
DeepSeek Harness
An open-source agent under the MIT licence in which every part, even the core, is a plug-in.
For an owner: Aimed at a developer, on a local machine, with DeepSeek's own model or others by plug-in. It is still a preview.
DeepSeek Harness is an open-source agent under the MIT licence. Every part of it, even the core, is a plug-in. It can be used from a terminal screen, run unattended or called from Python. It runs on Node (JavaScript). DSH Desktop is a separate community project.
- Can
- Loads plug-ins, opens its own web page on the machine, switches between model providers, DeepSeek's own among them, and runs as an official desktop preview, which arrived with v0.2 on 3 Oct 2026.
- Cannot
- Is still a preview: a stable release has not shipped yet.
Sources: MarkTechPost on DeepSeek Harness, v0.2 and official desktop apps (3 Oct 2026); DSH Desktop, community project under MIT (Read 6 Oct 2026); Vendor pages as read on 2 October 2026 (2 Oct 2026)
MiMo Code and MiMo Desktop
A terminal client under the MIT licence for coding and desktop tasks, with a desktop app beside it.
For an owner: Aimed at a developer who already works in a terminal. The client and the model are separate, so keys for other models can be used.
MiMo Code is a terminal client under the MIT licence, available since 15 Jun 2026. It is meant for coding and desktop tasks by a person who already works in a terminal. The MiMo-V2.6 weights were opened on 22 Sep 2026, and the MiMo Desktop app left early access the same day.
- Can
- Runs code tasks in a terminal and can use keys for DeepSeek, Kimi and GLM, so the client and the model are separate.
Sources: MiMo-V2.6 and MiMo Desktop (22 Sep 2026); Vendor pages as read on 2 October 2026 (2 Oct 2026)
Octop
One open-source program with a web dashboard, a command line, scheduled jobs and chat in seven messengers.
For an owner: Aimed at a household or a small team, on a local machine. It hands work to coding agents, so the coding agents do the work.
Tencent AI open-sourced Octop on 6 Oct 2026 under the MIT licence. One program gives a web dashboard, a command line, scheduled jobs and chat in seven messengers, for a household or a small team.
- Can
- Runs a multi-user workspace on a local machine (chat, browser, desktop) with a login per user and a workspace per agent, and hands work to Claude Code, Codex, Cursor, OpenCode and CodeBuddy. AgentTeams is in beta, mobile in closed beta.
- Cannot
- A plugin marketplace, managed agents and project workspaces are planned, not yet built.
Sources: Octop announcement by Tencent AI (6 Oct 2026); Octop repository (Read 6 Oct 2026); Octop releases, v1.0.2b6 (4 Oct 2026)
OpenClaw
A local gateway through which a person reaches an agent from messengers.
For an owner: Aimed at a person who runs the gateway, on a local machine. Its enterprise edition is in progress and is not yet 1.0.
OpenClaw is a local gateway. A person who runs the gateway reaches the agent through messengers. The Register reported on 30 Sep 2026 that firms were banning the consumer agent.
- Can
- Runs its own gateway and connects messengers to any model.
- Cannot
- Has no 1.0 enterprise edition yet: it is in progress, a management layer for many companies at once, started at OpenAI and donated to the OpenClaw Foundation.
Sources: OpenClaw enterprise edition (The Register) (30 Sep 2026)
OpenHands
A coding agent that you host yourself or run in the vendor's cloud, with any model.
For an owner: Aimed at developers. Agent Canvas gives engineers one screen to run and watch several coding agents.
OpenHands is a self-hosted coding agent that works with any model and can also run in the vendor's cloud; v1.25.0 was released on 5 Oct 2026.
- Can
- Agent Canvas drives OpenHands, Claude Code, Codex or any other compatible agent from one screen, so engineers have one place to run and watch several coding agents instead of a single command line.
Sources: OpenHands releases (Read 6 Oct 2026)
Osaurus
A native app under the MIT licence for coding on your own Mac or Windows computer.
For an owner: Aimed at one person on his own computer, with local models or a key. The Windows 11 version is a beta.
Osaurus is a native app under the MIT licence, so a repository can stay on the machine. It is aimed at a person on a Mac or Windows who runs it on his own computer.
- Can
- Does coding work with local models or a key on macOS, with a Windows 11 beta.
- Cannot
- Has no stable Windows 11 release yet; the beta was announced on 28 Sep 2026.
Sources: Osaurus repository (Read 6 Oct 2026); Osaurus blog (Read 6 Oct 2026)
Paperclip
An open-source org chart for agents under the MIT licence: each agent has one manager and budgets are set.
For an owner: Aimed at an operator who wires the seats himself, on a local machine. He brings in a coding agent for each seat.
Paperclip is an open-source org chart for agents under the MIT licence. Each agent has one manager, the CEO agent reports to the human, and budgets and regular check-ins are set for the agents. The operator brings in a coding agent. Its documentation describes the org structure, and the repository was updated on 6 Oct 2026.
- Can
- Runs an org of agents with budgets; an agent pauses when it reaches its budget cap.
Sources: Paperclip repository (Read 6 Oct 2026); Paperclip documentation (Read 6 Oct 2026)
SANKORD
You program your processes, not code. SANKORD speaks the language of your business, not the language of IT.
For an owner: For the one who decides in their business – with a team or still on your own – when the work no longer fits in one head. Not for a buyer who needs one layer across a whole company.
The agents work inside your walls: on your own servers or in your own cloud account. Models of your choice: open models on your hardware or cloud models through your own keys; you are not tied to one vendor. One agent prepares, another checks; what matters goes to you.
- Can
- Nothing irreversible happens without your yes: no payment, signature, deletion or publication. One command stops the whole system. Every step is written down. The record covers what the agents do and each yes they wait for, not how your people work.
Sources: SANKORD's own statements (October 2026)
Current as of October 2026
SANKORD and AIVIS
SANKORD is a project of the AIVIS company. Register on AIVIS and follow the project and its roadmap there.