What’s new.

• Agent Army: a company gets its own agents, each with a name, a role and its own CLI and model. Recruit the squad (Back, Front, QA, Docs, Growth) or hire your own.
• Three at once: up to three tasks run side by side, each in its own copy of the project. They land one at a time, only after the checks pass.
• Tasks that wait: a task can wait for the ones it needs, and each agent hears what the others delivered.
• Folders an agent owns: give an agent its folders, and its changes stay inside them.
• Review before landing, if you turn it on: the company's agent reads each change and sends back what isn't right.
• Backup: near its plan's limit, a task moves to the CLI you picked as its agent's backup, or waits for the plan to reset.
• A new look: flat panels, square corners and IBM Plex.

• The app's code is open at github.com/dhrive, under the Functional Source License: read it, change it, send a pull request.
• Feedback goes through your dhrive account, so the app carries no key of any service.
• Already have dhrive? Download it once more from dhrive.company: updates now come from the open repository, and they are automatic again from here.

• Checks on every delivery: after each task, dhrive runs the company's lint, typecheck, tests and build itself. A failing one goes back to the agent to fix; only a passing delivery is done.
• A standard for the code: every company's agent follows the same architecture and quality rules, and keeps the project's architecture written down.
• A model per company: pick the agent from the tasks bar whenever you want; waiting tasks follow the new pick.
• Environment: each company keeps its variables in your Keychain. Agents, terminals and Play get them, the published site gets only the public ones, and a removed one can be put back.
• Ask first is off by default and can be set when founding. Off, the agents run without stopping for approvals, and an MCP tools switch decides whether they reach your MCP servers.
• Feedback and the Discord community are a click away in the sidebar.

• Tools in a company's toolbar: connect Supabase, Stripe, Polar, PostHog, Resend, Sentry, Neon and Firecrawl. Each one installs its own CLI, signs you in through your browser and gives the agents its skills, so they can use it right away. Missing one? Suggest it from the same menu.
• Save to GitHub asks for a message: write it, or let the company's agent write it from what changed.
• A company's published website and starred address now show the same on every Mac and every install, read from the site host and the company's folder.

• A company can run on Ollama: pick it, and any model you have there, as the agent that founds and works for the company.
• Local models are now called Ollama in the sidebar.

Say what company should exist, and dhrive starts it and runs it: the product, its website and its documents, task after task, from the CLIs you already use.

The company

• One screen with everything side by side: tasks, website, documents and brief, with the chat beside them. The sidebar folds away.
• Tasks are suggested from the brief and worked one after another. Turn on Ask first to OK each one, or Auto mode to let it run for a few hours.
• The chat knows what the company did and turns what it suggests into tasks.
• Mission, roadmap, design guide and market research, written from what it built.
• Pick each step's model and effort.

The website

• Build checks it on your Mac first, and Fix it sends a failed build to a task. Publish puts it at its own address on dhrive.site.
• Every version is kept, and any earlier one can go live again.
• Connect your own domain: dhrive shows the two DNS records to set and what's still wrong until it's live. Star the address the company goes by.
• A preview sits under the versions; it opens the full site, the live one or the one running on your Mac.

Your data, kept

• Each company's tasks and timeline live in its own folder, so they travel with it.
• Put a company on GitHub in one click; after that, every finished task is backed up to its private repo.
• Deleting a company asks for its name, takes its site down and frees the name. Its folder goes to the Trash, and its conversations are put away rather than erased.

The maestro

• Each pane's copy now lives inside the project folder, hidden from git, so its CLI builds and runs the app where your project is.
• Bring together merges the panes' work into the project folder itself, on a branch of the session's own when you're on main, so the result is right where you work.
• Resolving a conflict or looking the result over opens in the project folder too.

The maestro

• Approving a plan on a project that isn't a git repo yet no longer stops the panes: the folder becomes a repo, its files go into a first local commit, and each pane gets its copy and starts coding.

The maestro

• Write a task once in Terminals and a CLI and model of your choice splits it into subtasks. It only reads the project.
• Pick the CLI and model for each subtask, read and edit it at full size, and point the maestro at the skills, MCP servers and plugins that matter.
• Approve and run: each subtask opens as a pane already at work, in a git worktree on its own branch. The split stays until you clear it.

Bring it together

• One click commits each pane's work and merges every branch onto one, without touching the branch you have checked out.
• A conflict is named with its files and opens in a pane to resolve. The result goes to a pull request in one more click.

Terminals live in your projects

• Terminals sessions sit in each project next to its conversations, each marked as what it is, and are kept between launches.
• Sessions the CLIs ran for dhrive (the maestro's, the panes') no longer show up as conversations, and any CLI session can be taken off the list.
• A CLI that's signed out says so in one line, with the command to sign back in.

A second way to work

• Switch to Terminals at the top of the sidebar and run up to six CLIs side by side: Claude Code, Codex and Antigravity, each as itself, each in a project of yours.
• Every pane shows when its CLI is working, when it's your turn, and the task it's on. A pane that needs you while dhrive is in the background sends a notification.

Never step on each other

• Give a pane a copy of its own and it works in a git worktree on its own branch, so two panes on one project never edit the same files.
• Turning the copy off removes it, unless it holds work that isn't committed yet. Then it stays where it is, and dhrive tells you where.

A fourth CLI

• Antigravity joins Claude Code, Codex and local models. Plan with Gemini and code with Claude, or any other mix: its models come straight from agy.

See it working

• A Browser tab beside the conversation shows your project's dev server. dhrive finds it by itself and reloads it when a code step ends.

Safer in parallel

• Conversations already run side by side. Now two code steps in the same project take turns, so one never edits over the other or undoes its work.

Your OK, your rules

• Nothing is edited without your OK: each plan waits for you to approve it, edit it, or reply to change it. One switch lets the code start right after the plan.
• When Claude wants to run something it isn't allowed to, it asks: allow once, always in this project, or deny.
• Your own instructions, like an AGENTS.md for every project, plus rules for a single conversation.

Stay in flow

• Conversations per project in the sidebar, with the sessions you had in the Claude app, the terminal or Codex beside them.
• ⌘K finds any conversation by anything that was said in it.
• Mention files and folders with @, attach screenshots, and open any file in a side panel.
• Watch each edit as Claude makes it. Requests queue while one runs, and you get a notification when it needs you.
• See how full each session's context is, and compact it before it compacts itself.
• Hover Claude Code or Codex to see your plan's limits, live. Click it to see every MCP server, skill and plugin it has connected.

Finish the job

• Every code step lists the files it changed, with the diff and an undo.
• Commit only that step's files, on a new branch, and open the pull request from the same place.

Your account

• Sign in once with your email or Google. You get 14 free days, no card.
• Then 1.500 $DHRIVE a month or 12.000 $DHRIVE a year. Your account on dhrive.company shows your days, your plan and where to manage it. Your CLIs keep their own plans.

dhrive is now a Mac app that gives every coding request two models: one plans, another writes the code, from the CLIs you already use.

• Plan with one CLI and model, code with another: Claude Code, Codex (including the copy inside the ChatGPT app) and open-weight models through Ollama.
• Local models are sized to your Mac before they download, and install in one click.
• Model lists come from each CLI, so new models show up without a dhrive update.
• Three modes: Chat to think it through, Plan for a plan then the code, Code to go straight to it.
• When the planner needs answers, it asks instead of guessing.
• Context carries over: each project picks up its CLI sessions, including ones you started in the Claude app, the terminal or Codex.