Your owncoding agent
One coding workspace for your models, skills and tools. Delegate work to parallel agents, follow their progress, and stay in control of the changes.
Desktop app and web workspace.
The agent runs on the machine that hosts it.
Example session: a workflow in the chat.
Example complete. Two child agents review code and tests in parallel, followed by a synthesis agent.
Work with
a whole crew
Delegate a piece of the task, or orchestrate the whole run. Keep the work in one conversation.
phase("Review");
const findings = await parallel([
() => agent("Review account code"),
() => agent("Review account tests")
]);
phase("Synthesis");
await agent(`Combine: ${findings}`);Subagents, coordinated
Start a fresh child agent or fork the current conversation. Run it in the foreground or background, with tools to message it, wait for it and stop it.
Dynamic workflows use a JavaScript orchestration script with agent(), parallel(), pipeline() and phase(). Claude Code dynamic-workflows meta blocks are supported. Runs appear live in the chat.
Plan before edits
In plan mode, the agent researches read-only. It then presents a plan for your approval before it edits anything.
The plan stays part of the conversation, so you can review the proposed work before it starts.
Hooks that fit
Use Claude Code compatible command hooks, configured in a hooks.json file.
SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStart and SubagentStop.
Skills for your work
Bring reusable instructions and supporting files in SKILL.md folders. Give the agent the procedures your project needs.
Keep project instructions alongside the code so they can be reviewed and versioned.
Tools and plugins
Extend the workspace with plugins and connect external tools through MCP. Choose the capabilities available to your agent.
Keep work moving
Track todos and use /goal to work toward an objective across turns. Background jobs and schedules support longer tasks.
Your models.
One workspace.
Choose a hosted model with Devvi credit, or configure a supported provider for your own installation. Model availability and limits depend on the connection you use.
Hosted models
DeepSeek V4.1 Flash and DeepSeek V4 Pro through your Devvi account. Usage appears in your account balance and history.
Devvi creditProvider connections
Configure supported model providers in your own installation and choose from the models available to that connection.
Your configurationChatGPT subscription
A personal connection is being tested for self-hosted use. Automatic fallback to Devvi credit is still in development.
Personal previewCredit, without
a clock
One devvi.sh account with prepaid credit: pay only for what you use, with no subscription, and keep whatever you do not spend.
Start with
$1
in credit once you confirm your email
Usage-based. Each request is charged from your balance at devvi.sh rates for the model you use, and your usage page lists every charge.
$5 to $500. Add credit by card through Stripe whenever you need more.
Your machine.
Clear boundaries.
Choose where the agent runs.
Your workspace. Your provider connection.
For hosted models, the provider credential stays on the Devvi server. For your own provider connection, credentials belong to your installation.
The agent works on the machine that hosts it. Model requests can include file contents and tool results needed for your task; the selected provider handles those requests.
Get the app
Windows release candidate 0.2.1-rc.3 is in validation. Public downloads will appear here after signing and installation checks.
Create your Devvi accountRun with care
devvi.sh is experimental developer-preview software. The agent executes model-written commands on the machine that hosts it.
Use the least privileges needed, preferably in a disposable VM or container. Keep backups and review proposed commands before allowing them to run. Permission controls do not guarantee isolation or prevent damage.