Repository navigation
Meevra Code 2.44.0
Meevra Code 2.44.0 - Public Preview
Meevra Code is an AI coding workspace for Windows. It works with the AI accounts you already have, keeps your work on your computer, and checks what its agents say they did. You open a project folder and describe what you want built, fixed or changed. A Lead agent plans and does the work with helper roles under your rules, and Meevra Code itself checks the result with real test runs before it calls anything done.
Download and install
- Download MeevraCode.exe below and run it. It installs Meevra Code on Windows 10 or Windows 11, 64-bit (x64).
- Read the license agreement and choose I Agree.
- Open Providers and connect at least one AI: an account you sign in to (Claude Code, ChatGPT or Grok Build), an API key, or a local model.
The preview installer is not code-signed yet, so Windows SmartScreen may show "Windows protected your PC". If you downloaded the installer from this page, choose More info, then Run anyway.
Already installed? You do not need this file. Meevra Code offers the update in Settings > Updates. An installed 2.42.3 takes this release as a small update (about 3 MB), and older versions take the full installer. Your projects, chats and settings are kept.
What's new in 2.44.0
- A terminal or a browser beside your work. Two buttons in the title bar open a pane next to the page you are on: Terminal (Ctrl+`) and Browser (Ctrl+Shift+B). Drag the pane's edge to make it wider or narrower.
- The browser pane. Tabs, an address bar that also searches, back, forward and reload. The pane has its own session that stays on your computer, so a site you sign in to there stays signed in. Forget sign-ins and site data clears it.
- The agent can work in the browser pane. Ask for it in a Code task, for example "open the preview in the browser pane and try the form". The agent gets its own Agent tab beside your tabs and opens pages, clicks, types, reads and takes screenshots there, while you watch. It works in the pane's session, so it is signed in where you signed in. Its approval request says so: "in your browser pane, where your sign-ins are".
- Thirteen built-in skills. They ship with Meevra Code and are on from the first start: video prompting, video editing with ffmpeg, Remotion, image generation, image editing, image upscaling, text in images, logo design in SVG, SVG quality, website design, model comparison, and two that choose the right image or video tool. The agent sees one line for each skill and reads a skill only when the task needs it. You find them in Marketplace > Skills with a Built in mark, and each can be turned off.
- Keep computer awake. A switch in the title bar's More menu keeps the computer from sleeping while a long task runs. It lasts until you turn it off or close the app.
- Current API prices. "What the API would charge" in Quota > Where it went uses the public list prices of 19 November 2026, for 217 models.
What came in 2.42.3
- A new Chat never starts on a model you removed. When you disconnect a connection in Providers, a new Chat no longer opens on a model of that connection. It uses the same model as Code until you choose one.
- A Chat you already used keeps its model. Its history stays as it was. The model is shown as not available until you choose another one.
What came in 2.42.2
- A window that stays blank at start recovers by itself. If the app's page has not arrived 20 seconds after start, Meevra Code loads it again. If it still has not arrived, Meevra Code makes the window again once, and after that shows the window as it is, so you are never left with an invisible app. Each of these steps is written to the app's log.
Everything in Meevra Code
Work
- Code. Describe a task once. A Lead plans, hands scoped work to helpers and integrates the result. The helper roles are Investigator (read only), Implementer (writes only inside the task's scope), Reviewer (read only), Diagnostician and Tester. A role's permissions are fixed: a prompt cannot give a read-only role the right to write.
- Chat. A conversation with a model that has no access to your files and runs no commands.
- Reply. Answer any message of a thread, with a quote, also while a task is running.
- Approval levels. Every step asks before anything that changes something. Balanced lets everyday development run and still asks for network access, installs, screen control and changes outside the workspace. Full access runs every action without asking.
- Checkpoints. Every step that changes files is saved, and you can rewind a session to any step.
- A branch per session (optional). A Code session can work in its own Git worktree, so your checkout stays untouched until you merge. Nothing is forced and nothing is pushed.
- Delivery. From a finished task to a pull request: Meevra Code pushes only the files the session's tasks changed, opens the pull request on GitHub, reads its checks and can send a failed check back to the agent. It asks you to confirm in its own window first. No model is used.
- Race. Several models do the same task, each in its own copy of the project. Meevra Code shows measured facts for each one and a judge's opinion, with the model names shown only after it. You take one version or combine them file by file, and the tests run again.
- Bots. Bots that tidy, summarise or digest your folders, with your approval.
- Night shift. Queue a task for the night. It runs on plan quota that would otherwise be lost before the reset, never on a route that costs money.
- Computer use. Agents can see the screen and use the apps you allow. It is off until you turn it on, the screen edges glow while it works, and Stop works at any moment.
- Terminal. Run approved shell commands inside the app, on its own page or in a pane beside your work.
- Browser pane. A browser with tabs beside your work. It keeps its sign-ins on your computer, and an agent can work in it when you ask.
- Keep computer awake. A switch that keeps the computer from sleeping during a long task.
Checks that use no model
- Proof it works. Before a task may end as done, Meevra Code runs your tests and your typecheck or build itself and shows the exact commands and their results.
- Test truth. It checks whether the agent's new tests would fail without the change. A test that passes either way is marked "Doesn't check the change".
- Claims checker. Every claim the agent makes at the end (tests passed, build OK, files changed, nothing left) is compared with Meevra Code's own records and shown as verified, not verified or contradicted.
- Proof strength. In a temporary copy of your project Meevra Code makes one small break at a time in the changed lines and shows which breaks your tests noticed.
- Error catcher. Errors from the commands, test runs and pages Meevra Code starts become cards mapped to your files, with Fix it.
- Stuck detector. A task that goes in circles is paused early, with a plain explanation, before it burns your quota.
- Promise. A running task shows how many turns and how much time tasks like it usually take, from your own record. A task that reaches twice its promise is held and asks you.
- Reflexes. A finished task that was a set of commands can be saved and run again with no model at all.
See what is happening
- Mission Control. Who works on what right now, the latest activity, and what waits for you.
- Flow. Every project at work, drawn live: each agent, its model and speed, the file it works on and its tokens.
- Activity. The audit trail of tasks, approvals, tests and provider events.
- Quota. Plan limits and usage exactly as providers report them, a forecast, who used what, and Where it went: every model turn by model, effort, task and day, with what the API would charge for the same tokens.
- Mobile. Pair your phone with a QR code and follow your tasks from it.
- TV & second screen (experimental). Extend or mirror Meevra Code to a TV or a second screen.
Memory and knowledge
- Brain. What Meevra Code remembers about your projects, chats and agents, stored on your computer. It learns from what really happened: a test command that ran and passed, a script that fails, what a finished task did. Agents get a small capsule of the memory that fits the task, as reference data and never as instructions.
- Vault. Notes and reference text, encrypted on your computer, that you add to a chat or a task yourself.
- Scout. A small, quick read-only helper that reads files for your agents, so the Lead's larger model does not spend its turns on reading.
Models and tools
- Agents. A model, a reasoning effort and a prompt for each role. The model is Fixed, Preferred (your ordered list) or Auto. Each agent can have its own budget, and team presets can be saved, exported and imported.
- Tools & MCP. The built-in tools and their rules, MCP servers, and a Marketplace of tools from the official MCP Registry.
- Skills. Thirteen built-in skills for video, images, SVG, websites and model comparison, on from the first start. You can add your own or install more from GitHub. An agent reads a skill only when the task needs it.
Supported AI connections
| Connection | How you connect |
|---|---|
| Claude Code | Sign in to Claude Code: Anthropic's own sign-in opens in its own window. Claude Code ships inside Meevra Code unmodified |
| ChatGPT | Sign in to ChatGPT: OpenAI's own sign-in opens in your browser |
| Grok Build (xAI) | xAI's own sign-in |
| Coding plans | The plan's own key: Z.ai GLM Coding Plan, MiniMax M Plan, BytePlus ModelArk Coding Plan, StepFun Step Plan |
| Anthropic API, OpenAI API, OpenRouter, Xiaomi MiMo | API key (OpenRouter also by sign-in) |
| Local models | The built-in engine with a .gguf file, Ollama, LM Studio, or an OpenAI-compatible server on your computer |
Meevra Code never sees your password: you sign in with the maker's own sign-in. One account per provider can be connected in the preview. Meevra Code includes no model access of its own, resells nothing and adds nothing to what your provider charges.
Your quota and your money
- Paid usage is Always ask by default. You can set a cap per task, per day or per month, or forbid paid APIs.
- A task that reaches a limit pauses and asks. It never moves to something that costs more on its own.
- A coding task is asked to finish in the fewest turns, and the turns after finished work run at a lower reasoning effort.
- The checks above use no model, so they cost no tokens.
Privacy
- No account with Meevra is needed, and no usage analytics are sent. Sharing anonymous usage totals is off until you turn it on.
- Your projects, chats, settings, memory and API keys stay on your computer. Your prompts and files go only to the provider you chose.
- The update check sends only the app's version and edition. Every update is verified before it is installed.
- The browser pane connects to the pages you open in it, and to the pages an agent opens there for your task. Those sites receive what any browser sends them. The pane's sign-ins and site data stay on your computer.
Known limits
- Outside a Git repository a task that ran a program ends as "Needs a look", not Done: without a repository Meevra Code cannot prove what the program changed.
- Sessions made before 2.42.1 show no "Answer to your reply" mark on their old messages.
- A Chat answer through Claude Code arrives whole when it is finished, not word by word.
- Skills are read in Code. Chat has no tools, so it reads no skill.
- The installer is not code-signed yet.
More
The full list of changes is in the changelog, and every feature is described in the README. Bug reports and ideas are welcome in Issues.
Meevra Code is free to use during the Public Preview under the End User License Agreement (EULA-en.txt). Open-source components are listed with their licenses in THIRD-PARTY-NOTICES.txt.