English · 简体中文

Meldwork builds the organization layer for AI agents.

Most tools help humans run more agents. Meldwork is built for a different job: make multi-agent work legible, accountable, and reusable.

The current preview is a local work cell for supported Agent CLIs. It keeps participants explicit, context scoped, collaboration bounded, run state inspectable, and human review in the loop. Selected Agents can now respond concurrently or move through independent proposals, peer negotiation, responsibility-based work, single-writer integration, and independent verification around one goal.

Download Meldwork V1.0.3 for Apple silicon macOS · Architecture · License

The missing layer

Agent supply is growing quickly. The organization around it is not.

Today, people still choose every Agent, move every piece of context, explain every role, reconcile conflicting answers, and carry the final accountability. Parallel managers improve throughput. Agent frameworks let developers program workflows. Single-Agent products improve one worker.

Meldwork is focused on the work between them:

• one shared goal and acceptance boundary;
• independent viewpoints instead of hidden routing;
• explicit authority, responsibility, handoffs, and stopping conditions;
• artifacts and evidence that survive Agent changes;
• a human-controlled adoption decision.

What works today

• Supported local Agent CLIs: detect, inspect, and invoke approved Agents without exposing unrestricted shell execution to the renderer.
• Persistent direct work: keep local conversation history, attachments, permissions, compatible native sessions, and sanitized execution details together.
• Concurrent responses: send one message to multiple selected Agents, freeze the same task context before execution, and publish their independent replies as one stable batch.
• Agent-negotiated collaboration: start with independent proposals, let the selected Agents challenge proposals and unanimously agree one responsibility graph, execute distinct work packages, integrate through one writer, and verify through different Agents.
• Inspectable execution: preserve bounded run events, warnings, completion states, checkpoints, and recovery actions without exposing credentials or private reasoning.
• Local control: keep Meldwork conversations and orchestration state on the computer; workspace writes remain opt-in.

Capability still depends on the selected Agent, installed version, authentication state, Provider, and declared tool or attachment support.

What Meldwork is building

The collaboration mechanism is not a hidden preset A → B → C assignment. The user selects the participants; those Agents expose the solution space, negotiate responsibilities, and execute inside Harness-enforced context, permission, budget, recovery, and evidence boundaries.

flowchart LR
  G["Goal + outcome contract"] --> P["Independent proposals"]
  P --> C["Visible challenge"]
  C --> B["Role claims + team bids"]
  B --> R["Responsibility contract"]
  R --> X["Bounded execution"]
  X --> E["Artifact + evidence"]
  E --> H["Human adoption"]
  H --> O["Outcome receipt"]
  O -. "Agent fit + team fit" .-> P

The selected-Agent proposal, challenge, responsibility, work, integration, and verification loop is available today. Automatic participant selection from a broader roster, remote Agents, enterprise governance, and outcome-driven reputation remain future direction.

The Harness

The Harness is the control plane that keeps conversations, context, adapters, streams, and durable state aligned in one workspace. It is what makes a run continuous across sessions, timeouts, recovery, and Agent changes.

Where Meldwork fits

This is a comparison of working models, not a feature-scorecard claiming Meldwork should replace every tool.

Approach	Examples	Best at	Main tradeoff	Meldwork's choice
One Agent CLI or AI IDE	Claude Code, Codex, Cursor, Windsurf	Deep work inside one Agent or product ecosystem	Moving to another Agent still requires a manual handoff	Preserve work across heterogeneous Agents
Parallel coding-Agent managers	Superset, Vibe Kanban, Nimbalyst, Conductor	Running multiple coding tasks and isolated workspaces in parallel	Primarily optimized for throughput and code delivery	Optimize proposal quality, responsibility, evidence, and adoption
Cloud Agent control planes	GitHub Agent HQ, Warp Oz, Devin, Factory, OpenHands Cloud	Dispatching and observing remote jobs, sandboxes, and Agent fleets	Work is organized around cloud or platform infrastructure	Start local-first and bring your existing Agents
Programmable Agent frameworks	LangGraph, CrewAI, AutoGen, Agents SDK, ADK	Building custom graphs, roles, routing, and automation	Developers must design and maintain the orchestration	Deliver an organization and acceptance experience to users
Multiple terminals and scripts	tmux, shell scripts, manual copy and paste	Maximum access to native Agent capabilities	The user becomes the router, context bus, and final judge	Make the work, boundaries, and decisions durable
Meldwork	Supported local Agent CLIs; remote supply is direction	Concurrent independent responses and selected-Agent proposal, negotiation, responsibility, evidence, and verification	Local single-user preview today; participants are still selected by the user	Build the organization layer for the Agent workforce

If one Agent and one work surface are enough, use the native tool. Use Meldwork when handoff, independent judgment, authority, evidence, and acceptance become part of the problem.

See the current work cell

Local Agent discovery	Direct multimodal work	Concurrent multi-Agent work
See which supported Agents are ready.	Keep prompts, files, permissions, and compatible sessions together.	Run concurrent replies or an Agent-negotiated collaboration with the participants you selected.

Connected Agents

Current integrations include:

Codex · Hermes · OpenClaw · WorkBuddy · Kimi Code · MiMo Code · Claude Code · Gemini CLI · OpenCode · Qwen Code

Specialized review:

OpenCodeReview

Availability depends on installation, version, authentication, Provider, and declared capabilities. Each Agent keeps its own tools, session behavior, and permission model.

Install

The current release-validated target is Apple silicon macOS.

1. Open the Meldwork V1.0.3 prerelease.
2. Download Meldwork-0.1.3-arm64.dmg.
3. Drag Meldwork into Applications and try to open it once.
4. This prerelease is ad-hoc signed, not signed with an Apple Developer ID, and not Apple-notarized. If macOS blocks it, open System Settings → Privacy & Security, find the Meldwork warning, choose Open Anyway, and confirm.
5. Connect a supported Agent CLI already installed on the computer, or configure an independent Provider profile where supported.

Use only artifacts downloaded from the official Release.

Build from source

Requires Node.js 22.12 or newer and npm.

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

Run the repository checks before distributing a build:

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

Trust boundaries

• Conversations and Meldwork orchestration state stay local.
• Model requests still follow the Agent and Provider you choose; local-first does not mean every model runs offline.
• Credentials use operating-system-backed secure storage where supported.
• Workspace writes are explicit and opt-in.
• Raw chain-of-thought, executable paths, secrets, and unrestricted command output are not exposed to the renderer.
• Cloud, Channel, automatic participant selection, enterprise governance, and an Outcome Network are not current release promises.

Read the architecture and Agent Connector contract for the public technical boundary.

License

Meldwork is source-available under the Meldwork Non-Commercial Source License. You may inspect, run, modify, and share it for non-commercial purposes. Commercial use requires prior written permission; see COMMERCIAL_USE.md.