Person memory
Keep preferences and stable user facts, such as a preferred language or answer format.
In-house Harness
Zleap-Agent
Organize tools and memory by workspace so the agent can focus on the current task.
WORKSPACE IS ALL AGENTS NEED
Keep the model focused on the current task
Zleap-Agent is our open-source harness for organizing agent tasks, context and tool use. For local and compatible API models, it first establishes the workspace, then supplies the tools, skills, memory and history needed for the current step.
Research runs in a search workspace; project file operations run in Cli. A task can move across workspaces, each with its own configuration and execution records.
Early previewApache-2.0
Workspace
Context organized around the task
Web + CLI
One runtime, two interfaces
Local + API
Connect local and compatible API models
01 / WORKSPACE IN ACTION
Enter the workspace each step needs
Follow a task from conversation to research and a finished document. Each step shows the tools, skills, memory and history provided in the current workspace.
Research public sources and write a product overview in the project, including citations.
TASK TRACE
CURRENT WORKSPACEMain
- Instructions
- Understand the goal, route work and summarize results
- Available tools
- Workspace selectionTask coordination
- Skills
- Task planning and summarization
- Memory
- User preference: lead with the conclusion
- Model
- Conversation model configured for this workspace
- History
- The user's goal and returned workspace results
1 / 6
Understand the goal
The user wants public sources researched and a short document written in the project. Main clarifies the goal and coordinates the task.
Illustrative task and workspace configuration. Shared system rules remain in place; the panel shows what varies by workspace. Autoplay stops after one pass.
02 / CONTEXT & MEMORY
Keep the rules, bring in the relevant context
Context includes system rules, workspace instructions, tools, memory and history. Zleap-Agent organizes these by how they change, separating stable requirements from information needed for the current turn.
01
Stable instructions
Persona, rules, workspace instructions and user preferences
02
Ongoing work
Retained conversation turns and the current event window
03
Retrieved for this turn
Recent messages and relevant past events or experience
Remember people, track events and retain experience
Event memory
Record facts and progress related to a user, task or workspace for later retrieval.
Experience memory
Retain reusable methods from completed work and retrieve them for relevant tasks.
03 / CAPABILITIES
Bring models, tools and execution together
The harness connects models to tools and relevant memory so they can carry out tasks. These capabilities are available in the public source; task and gateway integrations are still foundational.
One runtime for browser and terminal
Chat and manage workspaces in the Web UI, or run tasks through the CLI. Both use the same runtime.
Configure each workspace separately
Set instructions, models, tools and skills for the task. Engineering and research can use their own configurations.
Tools perform actions; skills explain how
Use built-in file, command and system tools, or add tools through MCP. Skills supply workflows, instructions and resources.
Choose a permission mode
Request approval or allow full access, controlling whether actions such as file writes and commands require confirmation.
Persist memory and execution records
PostgreSQL stores runtime data and memory, organized and queried by scopes such as user, task and workspace.
Foundations for ongoing tasks and channels
Task workers, scheduling and IM gateway code provide building blocks for ongoing work and external-channel integrations.
04 / USE CASES
Configure an agent around the work
Local-model applications
Configure workspaces for a local model, keeping available tools and context relevant to the current task.
Project work and research
Research sources, edit project files and create documents in separately configured workspaces, then return the results.
Internal systems and custom workflows
Use your model endpoints, MCP tools and permission settings to develop workspaces for engineering, operations or support.
05 / OPEN SOURCE
Run from source and configure it for your work
Zleap-Agent is open source under Apache-2.0 and currently an early preview for source review, local development and feedback. Prepare Node.js 20+, pnpm 9.x and PostgreSQL with pgvector, or use Docker to start the database.
Get the source
Clone the repository and install dependencies.
Start local services
Follow the docs to start the Web UI and database.
Connect models and tools
Configure a model, then set up workspaces for your tasks.
This page follows the public README and source. APIs, interface details and release flows may change.