Meldwork — Let Agents collaborate to move tasks forward

Set the goal. Let Agents work together. Meldwork brings your existing Agents into one working group to share context, discuss approaches, carry out work, and check each other's results — with less relaying, prompting, and stitching outputs together.

Meldwork is a local-first multi-Agent collaboration workspace for the Agent tools you already use on your Mac. Once you choose the participants, goal, context, and permissions, Agents in Auto Discussion can build on each other's work, request contributions or reviews, and advance the task within the authorized scope. Direct conversations and independent concurrent responses also support research, analysis, writing, planning, review, and implementation.

You do not need to direct every round, but you still set the goal, grant permissions, and make key decisions. Missing information, permission requests, or tradeoffs may require your input. Workspace writes are opt-in; collaboration history, run state, and deliverables stay local so you can revisit the work and continue it.

       

What Meldwork is for

Watch the product video on YouTube

Agent discovery

Auto discussion & run details

Direct multimodal conversation

Meldwork is useful when several Agents need to advance the same task and you want to spend less time passing context and coordinating next steps. Typical work includes:

• research and information synthesis across local Agent tools and selected knowledge sources;
• product, market, operational, and technical analysis;
• writing, planning, review, and structured document work;
• multimodal work with local documents, images, audio, video, PDFs, and code or configuration files;
• implementation in an authorized workspace, with other Agents checking the deliverables and requesting revisions.

How a task moves through Meldwork

1. Select the Agents that will work on the task together.
2. Scope the goal, working directory, context, attachments, Skills, knowledge sources, and permissions.
3. Collaborate in Auto Discussion: let Agents discuss approaches, carry out authorized work, and check results. Direct conversations and Concurrent Responses are also available.
4. Decide when the work needs information, permission, or a tradeoff resolved.
5. Check the result and request changes when more work is needed.

Collaboration modes

Mode	What happens	Best for
Direct	One selected Agent keeps its conversation and native session when supported.	Focused research, writing, analysis, or execution.
Concurrent Responses	Selected Agents receive the same frozen task snapshot and return independent responses in stable order.	Comparing perspectives before choosing a path.
Auto Discussion V4	Agents propose independently, discuss, request peer contributions, carry out authorized work, and check deliverables. A delivery owner judges progress and requests user decisions when needed. Native sessions continue when supported.	Multi-step collaboration with less manual coordination.

Participants are selected by the user. Automatic selection from a larger roster is outside the current preview.

Revisit the process and continue the work

Meldwork keeps task context, Agent replies, run state, and deliverables locally so you can see how work progressed, where it stalled, and what needs attention next. The Run Ledger preserves phase, participant, attempt, artifact, recovery, and Human Gate state. Diagnostic tool output and sensitive runtime details stay behind the Electron main-process boundary.

These records help you inspect collaboration; they do not establish that every claim has been fact-checked. Peer review does not guarantee correctness, and important deliverables still need acceptance checks appropriate to their use.

General Agent catalog

Meldwork detects and invokes compatible locally installed Agent commands. The catalog currently includes:

Codex	Hermes	OpenClaw	WorkBuddy	Zcode	Pi Agent	Kimi Code
MiMo Code	Claude Code	Gemini CLI	OpenCode	Qwen Code	OpenCodeReview	DeepSeek Harness

Approved Agent Connectors and custom executable Agents use the desktop connector paths. The desktop guide contains the current adapter, installation, provider, and capability matrix. The Agent Connector SDK guide explains supported packages, approval, upgrades, and conformance checks.

DeepSeek Harness (dsh) supports direct chats and group replies through ACP, with separate resumable sessions and workspace permissions. Its default route is ZGCI smart-router-v2; provide ZGCI_API_KEY through native credentials or the launch environment. In Settings → Agents → DeepSeek Harness, you can select an existing native DSH_HOME directory; detection, execution, and local Skills use that selection. The native DSH_AGENTS_HOME environment variable also controls the shared Skills directory. The icon comes from the official macOS application. An explicit ACP provider or model in the native profile or DSH home patch takes precedence over this default; credential environment variables declared there are passed to DSH.

WorkBuddy also supports selecting its existing native configuration directory in Settings → Agents. Its model list, execution environment, and local and plugin Skills use the selected directory.

OpenClaw also supports selecting its native state directory in Settings → Agents. Meldwork reads the selected Agent's model configuration from that directory and runs it in its guarded local runtime.

When switching an Agent's native configuration directory, Meldwork keeps separate local conversation sessions for each selected directory and restores them when switching back. Configuration changes are blocked while a run is active or awaiting recovery.

For Zcode, choose an existing .zcode/cli directory containing config.json. Zcode runs with that directory's parent user home, which also affects its home-relative session and other file paths; project configuration may still override user configuration.

For Qwen Code, choose an existing .qwen directory containing settings or credentials. Qwen uses its parent user home for native sessions, Skills and other home-relative files; project and system settings retain their native precedence.

For OpenCodeReview, choose an existing .opencodereview directory containing config.json. OCR runs with that directory's parent user home, including its native session and other home-relative paths.

Hermes uses its native config.yaml and Provider settings. When its model, fallback model, or provider entries declare credential variables with key_env or api_key_env, Meldwork reads those variables from the launch environment or login shell for Hermes; unrelated variables and unsafe process controls are excluded. Hermes may prepare its own dependencies when started with a new native data directory.

Context and local boundaries

Meldwork runs as a local Electron application. Conversations, group configuration, run records, and app-owned attachments stay in the local user data directory. The renderer receives only validated snapshots through a narrow preload API; executable paths, credentials, native session references, Skill paths, and unrestricted shell access remain in the main process.

Selected Agents may send prompts, attachments, or Skills to the model provider configured in their own CLI. Knowledge access is explicit and bounded: the current implementation supports local Obsidian retrieval and CLI-owned Feishu or DingTalk access modes. Local-first describes where Meldwork stores and coordinates work; it does not promise that an Agent's model provider is offline.

Run from source

Prerequisites: Node.js 22.12+ and npm.

npm --prefix frontend ci
npm --prefix desktop ci
npm --prefix desktop run dev

Before distributing a build:

npm --prefix frontend test
npm --prefix frontend run build
npm --prefix frontend run build:desktop
npm --prefix desktop test

Documentation

• Architecture and product boundary
• Desktop setup and Agent matrix
• Contributing
• Security policy

License

Meldwork is open source under the Apache License 2.0. Commercial use, private use, modification, and redistribution are permitted under the license terms. No separate commercial authorization or license fee is required. Apache-2.0 does not grant trademark rights; see COMMERCIAL_USE.md.