Developer tools · Coding agent

Coolwei Code

A coding agent that works inside one project folder at a time — reading files, editing them and running commands under a boundary you set, in a macOS application and at the command line. The whole system is hosted inside your own network.

Contact for trial → See it work

The application in its light theme: a new chat opened in a project, the project sidebar on the left, four starting suggestions in the centre, and a composer carrying the permission mode and the model.

The same new-chat view in the application's dark theme.

Coolwei Code is deployed inside your own environment. The application and the terminal agent talk to one endpoint — the model gateway your organisation already operates — and the models they offer are the catalogue your deployment publishes. Projects, chats, approvals, history and configuration are written to your own machines. No external connector, cloud task surface or hosted catalogue takes part.

macOSDesktop application, over the same agent as the terminal

2Entry points — the application and coolwei-code

3Permission modes, set per chat

0External connectors, cloud task surfaces or hosted catalogues

Watch the task run

A turn arrives as it is produced — the plan, the tool calls, the streamed answer — and stays readable after it settles. Nothing is summarised away.

A settled turn in the transcript: a plan card reading two of four tasks completed, the request underneath it, and the agent's written account of what changed with a before-and-after table and a code snippet. — The plan the agent is working through sits at the top of the turn; the composer footer carries the project, the change count with a link into Review, the permission mode and the model.

The same settled turn in the application's dark theme. — The plan the agent is working through sits at the top of the turn; the composer footer carries the project, the change count with a link into Review, the permission mode and the model.

An expanded activity group in the transcript, with a row each for reading a file, searching, listing files, editing two files and running the test suite, the command row showing its working directory and its output. — Every tool call keeps its own row — read, listed, searched, edited, ran a command — and a command row carries the directory it ran in, its output and its exit code.

The same expanded activity group in the application's dark theme. — Every tool call keeps its own row — read, listed, searched, edited, ran a command — and a command row carries the directory it ran in, its output and its exit code.

The Review panel open beside the transcript, headed Last turn with a change count, listing two changed files expanded into a syntax-highlighted unified diff. — Review holds the diff for the last turn or for the whole working tree, unified and syntax highlighted. The panel is read-only: nothing in it stages, reverts or commits.

The same Review panel in the application's dark theme. — Review holds the diff for the last turn or for the whole working tree, unified and syntax highlighted. The panel is read-only: nothing in it stages, reverts or commits.

What it does

Folder-scoped chats

Every chat belongs to a project folder you add to the sidebar, and the agent reads, edits and runs commands inside it.

The work in view

Tool-call rows, the plan and the streamed answer arrive as they are produced, then collapse into a group the turn keeps.

Approvals at the boundary

A command or a file write that would leave the sandbox stops the turn on a card carrying the request exactly as the agent sent it.

Review, read-only

The diff for the last turn or the whole working tree — unified, syntax highlighted, collapsible file by file, and never staged or committed for you.

Open the file the answer names

Files open as tabs beside Review, so a claim in the transcript can be checked against the source without leaving the chat.

Model and reasoning picker

A picker over the catalogue your deployment publishes, with the reasoning row shown for the models that offer one.

Resume in the terminal

coolwei-code picks up the same chats in the terminal that the application left open, in the same project folder.

Fully hostable inside your network

Both entry points are self-hosted deployments. Model traffic goes through your own gateway, the catalogue on offer is the one you publish, and everything the agent writes down stays on your machines.

The model picker open above the composer, listing the models the deployment publishes with a short description each, and a reasoning-effort row underneath. — The picker lists the catalogue your own deployment publishes. The reasoning row appears for the models that offer one, in the order the catalogue defines.

The same model picker in the application's dark theme. — The picker lists the catalogue your own deployment publishes. The reasoning row appears for the models that offer one, in the order the catalogue defines.

One endpoint, and it is yours

Model traffic follows the gateway configuration your deployment provides. There is no second destination to review.

Your catalogue, your names

The models and reasoning levels on offer come from the catalogue you publish, so the list matches what your platform team has approved.

Work stays on your machines

Projects, chats, approval history, transcripts and configuration are written locally, under a configuration home you control.

Nothing hosted in the loop

External connectors, cloud task surfaces and hosted catalogues are off in this build, and it does not call out for updates or announcements.

Three permission modes, one workspace boundary

The permission control sits in the composer and applies to the chat you are in. Two of the three modes share the same sandbox and differ only in who answers the approvals; the third widens it, and only after a confirmation that says so in as many words.

The permission picker open above the composer, headed Permissions for this chat and listing the three modes with a description each, the first one checked. — Every approval request offers four decisions: allow, allow for this session, decline, or cancel the turn. Anything still pending when the application quits is declined.

The same permission picker in the application's dark theme. — Every approval request offers four decisions: allow, allow for this session, decline, or cancel the turn. Anything still pending when the application quits is declined.

Ask for approval

Reads and edits inside the workspace. Asks before using the internet or going beyond the workspace boundary.

Approve for me

A reviewer agent decides the approvals at the same workspace boundary. Only who answers changes — the sandbox does not widen.

Full access

Edits files outside the workspace and reaches the internet without asking. Gated behind a confirmation and scoped to a single chat.

Light and dark, the same build

Both themes ship together and follow the system setting or an explicit choice. Every surface — transcript, Review, the file viewer, the pickers — is drawn in both.

The file viewer open as a tab beside the transcript in the light theme, showing the full path, the copy and open actions, and the numbered source lines. — Light

The same file viewer in the dark theme. — Dark

Coolwei Code is open for trials

Engineering, platform and security teams: contact the lab for a demonstration and a deployment plan for your own network.

Contact for trial →