Dev Agent

An IM-native coding agent

From conversationto code.

Describe the work. Choose direct implementation or clarify first. Dev Agent carries your request through implementation, checks, and review to a merge request — all from your chat.

For configured IM conversations and authorized repositories.

Dev Agent · Team chat
website

Static example · simple clarification shown below. All three modes are explained in the workflow section.

01 / Start with the outcome

You

/new Add a search filter to the results page.

Dev Agent

Let’s choose the repository and delivery branch, then a task mode.

You

example/website, targeting main. Keep the existing layout.

Choose a mode above to try a different path. These controls only change this fictional example.

02 / Make the details explicit

Dev Agent

Should search match titles only, or titles and descriptions? What should an empty search show?

You

Both, ignoring case. An empty search should show all results.

Dev Agent

Got it. I’ll also include a clear button and an accessible no-results message in the specification.

03 / Confirm, then give the go-ahead

Dev Agent

Specification summary
  • Search titles and descriptions, ignoring case.
  • Clear the search to restore all results.
  • Check filtering, empty results, and keyboard access.

You

I’ve read the full specification. This version is confirmed.

Dev Agent

Shall I begin implementation?

You

Yes, start building.

04 / Implement, check, and review

Dev Agent

I’m implementing the filter in the task workspace, then checking it against the agreed scope.
Implementation & checks

Filtering, clear behavior, and keyboard access.

Review · code quality

Code changes and repository conventions.

Review · requirement fit

The implementation against the task’s requirements.

Dev Agent

I’ll include the actual check results and review findings, including anything unresolved.

05 / A delivery you can build on

Dev Agent

The change is ready to inspect. You decide what gets merged.

Example merge request

Add a search filter

task branch → main

Code changes, check results, and review notes.
Not automatically merged or deployed.

You

Can we make the clear button more visible?

Dev Agent

Yes — I’ll continue in this task and update the merge request.

06 / Close with a clear recap

You

/end

Dev Agent

Task closed

Search filter delivered.

Total elapsed (incl. waiting)
18m 24s
Confirmation → latest delivery
8m 42s

Fictional elapsed times, including waits — not model runtime.

Dev Agent

Illustrated celebration of a completed coding task
Worth celebrating.

Fictional conversation · not a live session

Chat-native interaction Three modes. One task entry. Reviewable delivery Human-controlled merge

Agent Skills · Convention over configuration

Two ways in. Choose one.

Copy a prompt into your agent, or install reusable guidance with Skills CLI. Either is enough to get started.

Terminal · Install skills

One domain is enough: the Skills CLI finds our published skills and installs them for your coding agent.

Install from dev-agent.io

npx skills add https://dev-agent.io

Requires Node.js and npm. Installs guidance for your agent, not the Dev Agent service.

Give your agent the entry point

A prompt is a message that gives your agent context and instructions. Copy this into your agent, then describe your goal; it will read the command guide and help you choose the right task mode.

View prompt
Read https://dev-agent.io/.well-known/agent-skills/dev-agent-user-commands/SKILL.md. Help me choose a task mode and explain how to start and manage my Dev Agent task in the authorized IM conversation. Ask for my goal and any missing access or project details; do not assume access to the bot or repository. This prompt does not start or authorize implementation by itself.
Open the Markdown guide to read or copy manually ↗

How it works · index and manual handoff

curl -fsSL https://dev-agent.io/.well-known/agent-skills/index.json

Behind the command · well-known discovery

{
  "skills": [{
    "name": "dev-agent-user-commands",
    "url": "/.well-known/agent-skills/dev-agent-user-commands/SKILL.md"
  }]
}

The file format: Agent Skills ↗

Originally developed by Anthropic, Agent Skills defines the SKILL.md format: metadata plus task instructions. This is the format of the guide, not the website discovery protocol.

The index: a discovery proposal ↗

Cloudflare’s Agent Skills Discovery RFC proposes the well-known index, skill URLs, and digests. This site follows draft v0.2.0. It is a community proposal, not an IETF standard or a guarantee of client support.

The URL pattern: RFC 8615 ↗

The IETF’s RFC 8615 defines the general /.well-known/ URI mechanism. It does not define SKILL.md or this Agent Skills index.

The workflow

A clear path from request to delivery.

One /new entry, three ways to work. Every task keeps its repository, target branch, requirements, and reviewable delivery together.

Direct implementation

Describe the goal and let the agent decide routine details. Skip repeated clarification, specification writing, and a separate start confirmation. Checks and independent review stay in place.

Simple clarification

Resolve the essential gaps together. Read the specification and test plan, confirm that version, then decide when to start.

Deep clarification

Explore requirements, constraints, and trade-offs in more depth. Then confirm the specification and separately approve implementation.

  1. Choose where to work.

    Start with /new. Select an available repository and the branch your changes should target, then describe the outcome you need.

  2. Choose how to get started.

    Direct implementation starts from your request without a specification. Simple and deep clarification prepare a specification and test plan for you to confirm, followed by a separate go-ahead to build.

  3. Build, check, and review.

    Dev Agent works in a separate task workspace, runs checks, and independently reviews code quality and requirement fit. Results, unavailable checks, and unresolved findings are reported as they are.

  4. Make the result reviewable.

    Get a merge request against your selected branch. Inspect the changes, ask for refinements, and merge through your usual review process.

Built around the conversation

Keep the thread. Keep making progress.

Development rarely ends with the first draft. Continue in plain language, with the task and its context still in place.

01 / ITERATE

One task. Room to evolve.

Ask for another change in the same conversation. Before a merge, refine the existing request. After a confirmed merge, new work can become a new merge request without starting the discussion over.

02 / RESUME

Pause without losing the thread.

Check progress with /status, request a stop with /stop, and continue with /continue or a plain-language follow-up. Your mode, context, and work stay with the task; direct implementation stays spec-free.

03 / CONVERSE

Not every message is a task.

Ask questions, search public information, and read web pages in everyday chat. /new starts development; /end closes it. A thank-you or a question after delivery does not automatically trigger more code changes.

04 / CREATE

Simple visuals, without a code task.

Ask for an avatar, icon, or simple diagram in everyday chat when image delivery is enabled. These are programmatically drawn visuals, not photorealistic image generation. Adding images to a repository requires a separate confirmation.

Clear boundaries

The agent does the work.
You keep the decisions.

Choose how much to clarify. Repository access, image publication, and the final merge keep their own boundaries in every mode.

Work within the available scope.
Choose from the repositories configured for your conversation. The selected repository and target branch stay attached to the task.
A start that fits your mode.
Direct implementation begins from your request and handles routine details autonomously. Clarification modes require specification confirmation and a separate start decision. Missing essential information or permission can still block work in any mode.
See the result, including its limits.
A delivery includes the actual check and review status. Unresolved issues or unavailable checks are not presented as a clean pass.
A merge request is not a merge.
Changes are delivered on a task branch for review, not directly pushed or automatically merged into the target branch. Delivery does not automatically deploy your project.

The user guide

Already connected?
Start in your chat.

Send /help in your Dev Agent conversation, or use /new to begin a task. Your chat needs access to a configured repository; this website is the public product and command guide.

/new
Choose a repository, branch, and task mode.
/status
Inspect progress and existing delivery.
/stop
Request a stop; keep the task and its work.
/continue
Resume in context, or just say what to change.
/end
Safely end the task, or reset everyday chat.
/help
See the current commands in your conversation.