Workspace Is All Agents Need
English | 简体中文
Preview status
Zleap-Agent is an early preview. The public repository is intended for source review, local development, and feedback. APIs, UI details, packaging, and release flows may change before a stable release.
Zleap-Agent is a workspace-first agent harness for local and OpenAI-compatible models. It is built around one practical idea:
An agent should not see every tool, memory, rule, and previous message on every step. It should first know which workspace it is in, then receive only the context needed for that workspace.
Instead of treating context as one large prompt, Zleap-Agent separates the agent runtime into workspaces. Each workspace can have its own prompt, tools, skills, memory, model, and execution history.
This makes the system easier to reason about when running smaller local models, enterprise-local deployments, or workflows where permissions and data boundaries matter.
See how Zleap-Agent routes a task from Main into a focused workspace, keeps the tool context narrow, and brings the result back to the conversation.
zleap-agent-workspace-en-embed.mp4
- Workspace isolation for prompts, tools, skills, models, memory, and history.
- Web UI for chat, workspaces, assistants, models, tools, MCP servers, skills, memory, tasks, gateway configuration, and artifacts.
- CLI powered by the same runtime as the Web UI.
- PostgreSQL-backed memory and runtime persistence.
- Built-in file, command, system, and MCP tool support.
- Approval-required and full-access permission modes.
- Task worker and IM gateway foundations for long-running or external-channel workflows.
- OpenAI-compatible model provider support, with Anthropic provider code also present in the workspace packages.
A workspace is not just a tool group. It is an isolation boundary for the context the agent sees and the actions it can take.
Common examples:
Main: talks to the user, understands the goal, and routes work.Cli: reads and writes files, runs commands, and works inside a project.Web Search: searches and reads public web pages.- Custom workspaces: engineering, research, operations, finance, support, or any other domain-specific workbench.
Zleap-Agent treats context as a runtime layout:
Context = System Prompt + Workspace Prompt + Tools + Memory + History
The runtime avoids loading every available tool, memory, and trace into every turn. Workspaces keep the model focused on the current job.
Memory is partitioned instead of being stored in one generic bucket:
- Person memory: user preferences and stable user facts.
- Event memory: facts and state related to a user, task, or workspace.
- Experience memory: reusable methods learned from completed work.
PostgreSQL is used because memory participates in the agent loop and needs retrieval, isolation, auditability, and rollback.
Skills are reusable capability packages, usually centered around a SKILL.md
entry file. Tools are APIs; skills are workflows, instructions, examples, and
supporting resources that help an agent perform a class of work.
- Node.js 20+
- pnpm 9.x
- Docker Desktop, or a reachable PostgreSQL database with pgvector
git clone https://github.com/Zleap-AI/Zleap-Agent.git
cd Zleap-Agent
corepack enable
corepack prepare pnpm@9.15.0 --activate
pnpm installpnpm dev:webThis command starts the local development path:
- Loads environment from the repo root when present.
- Starts or connects to PostgreSQL.
- Builds the required workspace packages.
- Runs database migrations.
- Starts the Next.js Web UI.
Open:
http://localhost:3000
If you already have PostgreSQL available:
ZLEAP_DATABASE_URL=postgres://user:password@127.0.0.1:5432/zleap pnpm dev:webIn the Web UI, open settings and add an OpenAI-compatible model provider. For CLI experiments you can also use environment variables:
ZLEAP_MODEL_BASE_URL=https://api.example.com/v1
ZLEAP_MODEL_API_KEY=sk-...
ZLEAP_MODEL_NAME=qwen3.6-flashStart the CLI through the workspace package:
pnpm cliOr run a one-shot prompt:
ZLEAP_MODEL_BASE_URL=https://api.example.com/v1 \
ZLEAP_MODEL_API_KEY=sk-... \
ZLEAP_MODEL_NAME=qwen3.6-flash \
ZLEAP_DATABASE_URL=postgres://zleap:zleap@127.0.0.1:5433/zleap \
pnpm --filter @zleap-ai/cli start -- "Summarize this repository"Allow high-risk tools in one-shot mode:
pnpm --filter @zleap-ai/cli start -- --yes "Create a README draft in the current directory"pnpm dev:web # Web UI development loop
pnpm dev # Web UI + task worker + gateway development loop
pnpm dev:tasks # Task worker only
pnpm dev:gateway # IM gateway worker only
pnpm cli # Start the CLI
pnpm build # Build all packages
pnpm check # Build and type-check all packages
pnpm test # Run package tests| Variable | Description |
|---|---|
ZLEAP_DATABASE_URL |
PostgreSQL connection string for persistence, memory, skills, and tasks. |
ZLEAP_MODEL_BASE_URL |
OpenAI-compatible LLM base URL. |
ZLEAP_MODEL_API_KEY |
LLM API key. |
ZLEAP_MODEL_NAME |
Default LLM model name. |
ZLEAP_EMBED_BASE_URL |
Embedding provider base URL. |
ZLEAP_EMBED_API_KEY |
Embedding provider API key. |
ZLEAP_EMBED_MODEL |
Embedding model name. |
ZLEAP_EMBED_DIM |
Embedding vector dimension. |
ZLEAP_FILE_WORKSPACE_ROOT |
Default file workspace root when no project is selected. |
ZLEAP_WEB_SKILLS_ROOT |
Local skill directory scanned by the Web UI. |
assets/ public README assets
packages/ai model providers and model-call abstraction
packages/agent core agent runtime, workspace, skill, tool, and memory logic
packages/avatar inbound, scheduled, and web-chat run assembly
packages/cli terminal UI and CLI entrypoint
packages/core shared types and policy primitives
packages/gateway Feishu, WeChat, and long-connection gateway workers
packages/host local supervisor, Postgres bootstrap, install/update helpers
packages/runtime runtime-facing workspace and conversation APIs
packages/store PostgreSQL storage, migrations, and recall logic
packages/tasks scheduled task service and worker
packages/web Next.js Web UI
scripts/ source-development and package build helper scripts
The public GitHub repository keeps source code, README assets, and the helper scripts needed to run and build from source. Maintainer-only release packaging, signing, mirror-sync tooling, internal planning notes, local agent files, generated output, and secrets are kept out of the public tree.
Zleap-Agent is under active development. Areas still moving quickly include:
- workspace and model routing,
- memory extraction and retrieval quality,
- skill compatibility,
- gateway setup flows,
- desktop and release packaging,
- UI polish and end-to-end tests.
Feedback, issues, and focused contributions are welcome.
The license has not been finalized for the preview release. Please do not assume production redistribution rights until a license file is published.

