Project builder¶

Project builder is a scoped workspace for long-horizon tasks: software projects, research initiatives, reports, or any work that spans multiple sessions and needs persistent context, files, and a structured activity log.

What a project contains¶

Creating a project¶

Web UI (/projects)¶

1. Click New project.
2. Enter a name and goal. The goal is a natural-language statement of the project's purpose. Example: "Build a Python CLI tool that processes CSV files and generates summary charts."
3. Choose a project type (optional):
4. Software: adds a coding workspace and repository link
5. Research: adds a research run launcher
6. Report: adds a document editor
7. Blank: no type-specific UI
8. Click Create project.

Via API¶

POST /api/projects
{
  "name": "CSV chart tool",
  "goal": "Build a Python CLI tool that processes CSV files and generates summary charts.",
  "type": "software"
}

Working in a project¶

All conversations started from the project page are automatically scoped to the project. The agent:

• Reads the project goal at the start of each turn.
• Indexes and retrieves project files and notes for context.
• Logs every tool call and file change to the activity log.
• Creates sub-tasks automatically when it plans multi-step work.

Uploading files¶

Drag files onto the project page or use the file panel. Uploaded files are immediately available to the agent as context.

Linking a repository¶

Software projects can link a GitHub repository:

1. Open the project, go to Settings > Repository.
2. Authenticate with GitHub.
3. Select or create a repository.

Once linked, the self-coding agent can open branches, commit code, and create pull requests within the project.

Sub-tasks¶

The agent decomposes work into sub-tasks automatically. You can also add them manually:

POST /api/projects/{project_id}/tasks
{"title": "Implement CSV parser", "status": "pending"}

Sub-tasks appear in the project's task list with status: pending, in_progress, done, blocked.

Memory scope¶

Each project has its own vector memory collection. Documents uploaded or indexed within a project are only searched in the context of that project (unless you explicitly search the global memory).

This prevents cross-project contamination: a file in your "client A" project won't surface in your "client B" project's searches.

Activity log¶

Every agent action is written to the activity log with:

• Timestamp
• User and session
• Tool called
• Input/output summary (not full content)
• Status (success/failure)

View the log in Project > Activity or query it:

GET /api/projects/{project_id}/activity?limit=50&tool=write_file

Project templates¶

Save a project as a template to reuse its structure:

POST /api/projects/{project_id}/save-as-template
{"name": "Python CLI project"}

Create a new project from a template:

POST /api/projects
{"name": "New CLI tool", "template_id": "python-cli-project"}

Archiving and export¶

Archive a completed project to hide it from the active list:

POST /api/projects/{project_id}/archive

Export a project (files, notes, conversation history, activity log) as a ZIP:

GET /api/projects/{project_id}/export

API¶

Related¶

• Self-coding agent
• Memory and RAG
• Playbooks
• Agent teams