MICRO-MULTI

Micro-Multi:
A next-generation
multi-agent workbench

The lead splits the job. You approve the assignments and choose each member's model. Compatible with DeepSeek Harness plugins.

Windows · Linux · English and Chinese · Apache-2.0

● ● ●Example task
PROJECT TASK

Add pagination to the order API and test boundary cases.

M
Lead agentPrepare the plan and review the changes
Review the plan
01
ImplementationUpdate the endpoint
02
TestingCover boundary cases
Approve → Schedule → Combinedsh · Shared tools

Collaboration needs a scheduler

01

Split work by responsibility

The lead proposes assignments for implementation, tests and docs. Review the members, their files and models before execution.

02

Run independent work together

Independent members work concurrently. Tasks with dependencies wait for predecessor reports. You choose the concurrency limit.

03

Coordinate shared files

File ownership and path locks coordinate writes. The scheduler reuses running assignments when the lead dispatches them again.

04

Share dsh tools with the team

Enabled dsh plugins supply tools to the lead and members. Plugin settings and interface components live in the same workspace.

How team execution works →

Compatible with DeepSeek Harness plugins

Install DeepSeek Harness plugins in Micro-Multi and share their tools across the team. Manage plugin settings in the app, and add Skills or MCP services when you need them.

See supported plugins and setup →
Plugin toolsPlugin settingsSkillsMCP

When to use multiple agents

Use one agent for a small edit. Use multiple agents when a job includes implementation, tests and documentation, or spans several parts of a project.

ProblemA single loop / uncoordinated agentsMicro-Multi
WaitingImplementation, tests and docs run in sequence.Independent tasks run together; dependent tasks wait.
Shared filesConcurrent edits need manual coordination.Declared ownership and path locks coordinate writes.
Model choiceOne model often handles every role.Choose individual member models in the plan.
RetriesAn assignment may be dispatched again.Reuse running or completed assignments.

More agents can cost more. This is an execution comparison, not a speed or price benchmark. Use only the members the job needs.

Design choices and tradeoffs →

Start with a model and a project

You supply the model endpoint and credentials. The app handles the conversation and workspace; your project keeps its own build tools and dependencies.

01

Connect a model

In Settings → Models, add the API base URL, model ID and key. Test the connection before starting.

02

Open a local folder

Add your project and choose whether to link the original directory. Check the path before confirming.

03

Review the assignments

Choose team mode, describe the job, and inspect the proposed members, files, dependencies and models.

04

Confirm and follow the work

Confirm the plan in chat. Follow each member, then check the combined diff and test output.

Read the workspace guide →

Technology

An Electron desktop shell runs a local Python backend. Node.js hosts the agent and plugin runtimes, while the interface shows their work and records.

Desktop

Electron

Folder selection and desktop integration.

Backend

Python · FastAPI

Local APIs, conversations and task execution.

Interface

HTML · CSS · JavaScript

The main workspace; React for native plugin components.

Agent and plugin runtime

Node.js · Cordis · dsh

Native tools, services and plugin contributions.

Storage

SQLite · keyring

Local records and OS-backed model credentials.

Connections

HTTP · SSE · WebSocket

API calls, streamed replies and plugin subscriptions.

Read the architecture guide →

Install Micro-Multi

Choose a package for your system. Python and Node.js are included; install Git for repository work and any external tools your extensions require.

SHA256SUMS · Installation and Linux requirements · Run from source

Help and contributions

Check the troubleshooting guide when something fails. You can report bugs, discuss ideas or contribute code and documentation on GitHub.