AI coding agent · Free for Windows

Your code. Your models. Your machine.

Orynr Code is a free AI coding agent that works on your own computer, with the models you choose, in the cloud or on your own GPU. A team of specialist agents plans, builds, checks and fixes the work, safely sandboxed, and built so a job never silently dies halfway.

Version 2.0.0 · Windows 10 and 11, 64-bit · 82 MB · No admin rights needed · Free, including for commercial work

One command Ready

Type ora and say what you want.

Open Orynr Code from the Start menu, or type ora in any Command Prompt or PowerShell. ora C:\my\project opens a folder directly.

Describe the job in plain words. The Agent works out what you actually need, hands each piece to the right specialist, reviews their work and tells you what it did.

  • Nothing else to install. No Python, no Node.js.
  • A place for every project. On your first chat it suggests a name and makes the folder in Documents\Orynr_Projects, or works right in a folder you already have.
ora · shop-api
PS C:\> ora C:\dev\shop-api

 Add rate limiting to the login endpoint, with tests.

● Agent  Limit failed logins per account; keep the API the same.
   planner       3 steps: middleware, config, tests
   builder       edit src/auth/login.py        +42 
                 run  pytest tests/auth        14 passed
   risk checker  low: no schema or dependency changes
   verifier      checked the project against your ask 

● Done. Five tries a minute per account, three new tests.

An illustration of a session, not a screenshot.

What makes it different

A whole engineering team in your terminal.

Not a chatbot that hands you snippets. A team that finishes the job, on your machine, with the model you pick.

A team, not a chatbot Ready

The Agent works out what you need, then hands each piece to a specialist: researcher, planner, builder, code auditor, risk checker. It reviews their work and reports back. If no specialist fits, it builds a new one on the spot.

Any model, including yours Ready

75+ cloud providers, or your own model server on this PC or your network: llama.cpp, vLLM, LM Studio, Ollama. Type its address and Orynr asks what is really loaded, so the list is never stale.

Never dies, never goes silent Ready

Long installs and builds get up to 30 minutes. If one still runs out of time, the agent checks what actually happened before trying again. After six failed tries it tells you plainly.

Talk to it while it works Ready

Live Steering: type a new message mid-task and it stops, reads it and changes course at once. No waiting for it to finish the wrong thing.

A real sandbox Ready

Run a project in a locked-down Docker container that can reach the internet and nothing else: not your PC, not your home network. Or run it locally. Switch any time and Orynr moves the files.

It checks its own work Ready

A verifier looks at the real project before a job is called done, and sends it back if something is missing. After every edit the language checker runs, and new errors get fixed.

Private by design Ready

No telemetry, no account, nothing sent to Orynr. Your API keys are stored encrypted with AES-256, locked to your Windows account.

Projects that remember Ready

Each project keeps its own long-term memory and a map of the codebase, and starts the next session with a short “since last time” brief.

Plug in any tool Ready

Connect any MCP server, remote by web address or a program on your PC. Orynr tests it first and keeps its key encrypted. Or ask the Agent to build a new one for you.

The agent team Ready

You ask once. The right specialists do the work.

The default Agent, Orynr, is a friendly orchestrator. It turns your request into pieces, sends each to the specialist best at it, runs them in parallel where it can, and reviews what comes back before a verifier checks the real project.

When nothing on the team fits, it creates a new specialist, or even a new tool server, on the fly.

You“Add rate limiting to login, with tests.”
Agent (Orynr)works out what you actually need
ResearcherPlannerBuilder Code auditorRisk checker+ a new one, if needed
Verifierchecks the real project against what you asked
Back to youwhat it did, and what it found

Or pick the mode yourself Ready

Fourteen ways to work, from look-but-don’t-touch to keep-going-until-it’s-done.

Agent (Orynr) · default

Understands what you want, delegates to specialists and reviews their work. Can create new specialists and tool servers.

Build

Does the work directly: edits files, runs commands.

Autonomous

Keeps going on its own until the task is done and verified, working through its own to-do list.

Ralph

A relentless, spec-driven finisher that drives an objective to completion, with a hard step limit.

Plan

Look but don’t touch. Produces a plan without editing anything.

Chat

A switchboard that routes your request to the right specialists.

Conversation

A plain chat companion that explains and writes code as text, touching nothing.

Research

Investigates the web and your code in parallel, and reports with sources.

Thinking

Reasons through hard problems in parallel, from your files.

Understanding

Turns a vague request into a clear, structured brief.

Mapping

Maps how a codebase fits together: modules, calls, data flow.

Code Analyzer

A read-only audit for bugs, smells, security and performance risks.

Risk Analyzer

Rates how dangerous a proposed change is before you run it.

Anomaly Reviewer

Scans for leaked secrets, hard-coded paths, debug leftovers and over-complex code.

Behind the scenes, helper agents run work in parallel, a verifier checks results and an MCP builder creates new tool servers. Add your own agents in the settings, or let Orynr create them.

Models

Bring any model. Keep every file.

Use a cloud provider, your own GPU, or both, and switch model or mode in the middle of a chat.

  • Your own server comes first. Type an address like 192.168.1.50:8080, a name and an optional key. Orynr checks it and tells you plainly if the address, port or key is wrong.
  • Live model discovery. It asks your server which model is loaded, with its size and context length.
  • Think mode. For local models that support it, turn on think or think harder, and watch the thinking stream live.
  • Models without tool support still work. Orynr teaches them to write tool calls as text and reads them back reliably.
  • Busy servers, handled. A GPU busy with someone else’s job is waited on, with a live “the model server is busy” line. Requests queue instead of colliding, and you can stop any time.
  • Images, costs, sign-in. Paste screenshots for vision models, see cost and tokens per request, and sign in with OAuth where a provider supports it.

Your own server Ready

This PC or your networkllama.cppvLLM LM StudioOllamaAny OpenAI-compatible server

75+ cloud providers Ready

OpenAIAnthropicGoogleDeepSeekxAI OpenRouterGroqMistralGitHub Copilot AzureAmazon BedrockCerebrasFireworks Perplexityand more

A built-in catalogue of thousands of models, with names, prices and context limits. Cloud models use your own account and keys.

Reliability Ready

The coding agent that finishes the job.

Long jobs are where agents usually fall over: a build times out, a model stops answering, and the work quietly dies. Orynr is built so that doesn’t happen.

01

Nothing dies silently

Every wait, whether a command, a helper or the model, may take up to 30 minutes. Then the agent checks the real state and tries again. After six failed tries it tells you what it was doing and what it found.

02

“Timed out” is not “failed”

A long install that ran out of time is checked first (is Python installed now?) before anything is repeated.

03

Retries you can see

A model that doesn’t answer is retried on screen, with growing pauses and no repeated text. A wrong API key is reported at once.

04

Nothing gets lost

A helper stopped part-way hands back what it did and found. One waiting for your “Allow” waits as long as you need.

05

Verified, then done

Before a job is declared done, the real project is checked against what you asked for. Each edit is checked by the language server, and new errors are fixed right away.

06

Long chats never overflow

Older parts of a conversation are condensed automatically as it nears the model’s limit.

ora · data-pipeline
builder   run   pip install -r requirements.txt
          
          check is the environment complete?
          yes: 118 packages installed
          carrying on, nothing to repeat

model     
          Esc to stop
          trying again, attempt 2 of 6
          answering

verifier  3 of 3 requirements met 

An illustration, not a screenshot.

The sandbox

Three ways to run a project. Switch any time.

Every project chooses how its commands run. Type /sandbox to change it: Orynr moves the files between modes, never overwrites anything, and rolls back if a move fails.

Docker Ready

Only the project’s own folder is given to a container, and every command runs inside it. The container reaches the internet and nothing else: not your PC, its other folders or other machines on your network, enforced by a firewall inside the container. Commands run as an ordinary user with special privileges removed.

Docker shared local Ready

The same locked-down container, working directly on your project folder. Nothing from the project runs on your PC.

Local Ready

Commands run straight on your PC. The fastest, and the default when Docker isn’t installed.

Good to know The Docker modes need Docker Desktop; without it, projects run in Local mode. You can copy any folder into a sandboxed project, and the AI itself can never reach your other folders. Quick snippets of Python, JavaScript, Ruby, PHP or Lua run in a separate WebAssembly sandbox that can’t touch your machine at all.

Safety and control Ready

It asks. You decide.

  • Permission prompts before risky actions: allow once, allow always, or reject with feedback. “Allow always” shows exactly what it covers, and never lets a chained or hidden command slip through.
  • Undo and rewind. Every turn snapshots the project, so you can roll your files back to any earlier point, even without git.
  • Stop anything mid-answer with Esc.
  • Extra care for dangerous commands. Optionally require confirmation for things like force-push or rm -rf.
  • Project trust. A project’s own tools, skills and settings load only once you trust it.
  • Off limits. System folders, whole drives and your home folder can never be adopted as a project.

Private by design Ready

  • Where it runs: on your computer.
  • Telemetry: none. No analytics, no account.
  • Your code and messages: go only to the model provider you choose, never to Orynr.
  • API keys: encrypted at rest with AES-256-GCM, the key protected by your Windows account.
  • Offline: it needs a model it can reach. A server on your own network counts; cloud models need the internet.

Some features download what they need, such as a code-intelligence server or a container image. Those downloads carry none of your code or messages.

Extensible

Plug in any tool, local or remote.

Orynr speaks MCP, the open standard for tool servers. Add one in a few steps from the MCP screen, or just ask the Agent in chat.

A remote server Ready

Paste its web address and, if it needs one, an API key or token. The key is stored encrypted, never as plain text in a settings file.

A program on this PC Ready

A command such as npx -y @modelcontextprotocol/server-github or a path to a server file, plus any key it needs, like GITHUB_TOKEN, also encrypted.

Or just ask Ready

Say “connect the MCP server at https://…” and the Agent connects it. It can even build a brand-new tool server, register it and check that it connects.

Tested before it’s saved Orynr connects to every server first, shows which tools it offers, and explains plainly if the address, key or command is wrong. A server on your own PC or home network is blocked by default and allowed with one confirmation. You can also drop a server into a project’s MCP folder, and add skills (including ones written for Claude Code), plugins and custom agents.

Remote servers need nothing extra. npx servers need Node.js, and servers written in Python, including the ones Orynr builds for you, need Python. Ready-made tool servers for Excel, PowerPoint, documents, databases, code intelligence, system monitoring and more are Coming soon to the installer.

And the rest Ready

Works where you work.

Terminal UIThemes, a command palette (Ctrl+P), keyboard shortcuts, live tool progress and helper views.
EditorsDrive the agent from editors that speak the Agent Client Protocol, such as Zed.
Command line18 non-interactive commands (run, session, models, provider, mcp, stats, export and more) for scripts and automation.
HTTP APIThe engine runs as a local server that the terminal UI, and your own tools, talk to.
Code intelligenceBy symbol name: hover for types and docs, go to definition, find all references, safe rename across the whole project, diagnostics, formatting and completions.
ToolsRead, write and edit files (including a “fuzzy patch” that still applies when a file has drifted), run commands, search by name and content, search the web with no key needed, fetch pages, run and preview a dev server, and ask you multiple-choice questions.
MemoryLong-term project memory, a saved codebase map, a “since last session” brief, and your ORYNR.md, CLAUDE.md or CONTEXT.md instructions.
SessionsFull history per project. Share a conversation as a read-only page served by your own machine, and revoke it. Export and import as JSON. Sync sessions peer-to-peer between machines.
UsageCost and token tracking per request, with usage statistics.

Download

Orynr Code 2.0.0 Ready

Free to use, including for commercial work.

Download for Windows · 82 MB

  • File: orynr_code_setup.exe
  • Size: 82.4 MB (82,418,736 bytes)
  • Works on: Windows 10 and 11, 64-bit
  • Installs: for you only, no admin rights. Clean uninstall, which also removes it from PATH.
  • SHA-256: A84747F8A36395813889EDEE9BF82C228B922FF114E7AAA5E7CFB9D9709ED429
  • Signed by: Orynr LLC, certificate thumbprint 5B76195D3732254A8675BB882C51A4CB08C2F0A7
  • Mac and Linux: Coming soon

Licence agreement (EULA) · Third-party notices

Windows may warn you the first time The installer is signed by Orynr LLC with our own certificate, which Windows doesn’t recognise yet, so SmartScreen says “unknown publisher”. Choose More info, then Run anyway.

To be sure your copy is ours, check its fingerprint first. In PowerShell, in the folder you saved it to, this should print True:

(Get-FileHash .\orynr_code_setup.exe).Hash -eq 'A84747F8A36395813889EDEE9BF82C228B922FF114E7AAA5E7CFB9D9709ED429'

Getting started

  1. Run the Setup. A normal Windows installer, with Orynr Code in your Start menu.
  2. Open it from the Start menu or desktop, or type ora in any terminal.
  3. Add a model: a cloud provider’s key, or the address of your own model server.
  4. Say what you want. The Agent takes it from there.

Optional: Docker Desktop, for sandboxed projects.

Orynr Code is published by Orynr LLC, Florida, USA. Its terminal interface is based on the open-source OpenCode project (MIT licence); the engine behind it is Orynr’s own. Questions: info@orynr.com.