What is Meldwork and what does it do?

Meldwork is a local-first General Agent workspace for macOS. It lets you coordinate local Agent tools for research, analysis, writing, planning, review, and implementation. Meldwork freezes one task snapshot, sends it to every selected Agent, captures independent findings as evidence, and requires your approval (Human Gate) before any workspace changes are written.

How is Meldwork different from tmux or zellij for running multiple AI agents?

tmux and zellij are terminal multiplexers that show agent output side by side but require you to manually copy context between agents. Meldwork freezes one task snapshot and delivers it to all selected agents automatically, then captures their responses as structured evidence (Finding → Evidence → Decision → Disposition) with a human adoption gate before any file changes. It also preserves session state across restarts, so you can resume interrupted multi-agent work.

Which General Agents does Meldwork support?

Twelve agents today — Codex, Claude Code, Gemini CLI, Qwen Code, Kimi Code, MiMo Code, OpenCode, OpenCodeReview, Hermes, Pi Agent, OpenClaw, and WorkBuddy. Approved Agent Connectors can be added through the Agent Connector SDK; custom executables use the desktop custom-Agent path.

Can Meldwork run Codex and Claude Code at the same time on the same task?

Yes. In Concurrent Responses mode, Meldwork sends the same frozen task snapshot to Codex, Claude Code, and any other selected agents simultaneously. Each agent returns an independent response, and you compare their findings side by side before deciding which result to adopt. In Auto Discussion V4 mode, agents can also propose, challenge each other's proposals, negotiate responsibilities, and verify results across multiple rounds.

What does Auto Discussion V4 actually do?

Selected Agents independently propose, then critique with evidence, negotiate responsibilities, execute dependency-aware work, and verify the result. Every round is preserved — you can read the full Case later and adopt only the parts you approve.

How is Meldwork different from Claude Code Agent Teams or CrewAI?

Claude Code Agent Teams only works within Claude Code — you cannot mix Codex, Hermes, or OpenCode. CrewAI is a Python framework you build yourself. Meldwork is a ready-to-use desktop app that orchestrates 12+ heterogeneous agent CLIs from different vendors in one workspace, with evidence trails, a human adoption gate, and no orchestration code to write.

Is Meldwork free? What's the license?

Source-available under the Meldwork Non-Commercial Source License 1.0. You can inspect, run, modify, and share it for non-commercial purposes. Commercial use requires prior written permission.

Does Meldwork work on Intel macOS, Windows, or Linux?

The current V1.0.5 preview targets Apple silicon macOS (ad-hoc signed, not notarized — open Open Anyway in System Settings → Privacy & Security on first launch). Cross-platform builds are on the roadmap; the app is Electron so the same source compiles everywhere once those targets land.

Meldwork — A Local-First Workspace for General Agents

DETECTED ON YOUR MACHINE

Works with the agents you already run.

Meldwork detects a compatible local Agent command and exposes its readiness, capabilities, and session behavior. Approved Agent Connectors can be added through the Agent Connector SDK.

THE SHORT ANSWER

What is Meldwork?

Meldwork is a local-first General Agent workspace for research, analysis, writing, planning, review, and implementation. It freezes task context, captures independent findings as evidence, and puts a human adoption gate before any workspace change.

Supported agents 12+

Codex, Claude Code, Gemini CLI, Qwen Code, Kimi Code, MiMo Code, OpenCode, OpenCodeReview, Hermes, Pi Agent, OpenClaw, WorkBuddy — detected natively on your machine.

Modes 3

Direct · Concurrent Responses · Auto Discussion V4

Review record

Finding → Evidence → Decision → Disposition

Attached to every result — dissent stays visible, never hidden behind a manager.

Runs on

macOS · Apple silicon · Electron

100% local-first work cell

THE WORK CELL

Four steps from task to adopted result.

01

Select

Choose the local Agents and participants. Participants are always selected by you — never auto-assigned from a roster.

02

Scope

Bound the goal, working directory, context, and permissions. Every run starts from a frozen task snapshot.

03

Run

Direct, Concurrent Responses, or Auto Discussion V4 — from a single focused session to negotiated multi-round work.

04

Review & adopt

Inspect findings, evidence, and any Human Gate before the workspace changes. Nothing writes itself.

COLLABORATION MODES

Three ways agents work together.

Same Case, same evidence trail — different depth of agreement. Pick the mode that matches how much review the decision deserves.

>

Direct

One selected Agent keeps its conversation and native session when supported.

Best for focused work with one Agent.

*

Concurrent Responses

Selected Agents receive the same frozen task snapshot and return independent replies in stable order.

Best for comparing approaches before choosing one.

#

Auto Discussion V4

Agents propose, challenge, negotiate responsibilities, execute dependency-aware work, then verify the result.

Best for multi-round work with explicit responsibility.

WHAT YOU CONTROL

Four guarantees that hold the work together.

These four invariants are what make multi-agent output reviewable instead of just decorative.

Human Gate

Workspace writes are opt-in. Nothing changes until you adopt the result — agents don't.

Opt-in workflow controls

Frozen Context

Every run starts from a snapshot. All selected Agents see the same brief, in stable order.

Case-scoped, reproducible

Evidence Trail

Finding → Evidence → Decision → Disposition, attached to each result. Dissent stays visible.

No hidden manager

Local-first Boundary

Electron work cell on your machine. No remote Agent fleet. Conversations stay local.

Your CLI, your data

INSIDE A REAL RUN

Launch Readiness Council — three rounds, one gate.

Below: the actual evidence from a V1.0.3 session where Codex, Hermes, and OpenCodeReview reviewed a private-beta release decision. Scroll or wait — the timeline advances on its own.

Auto Discussion V4 Launch Readiness Council

Round 1 · Propose

R1 · 05:34

CodexProposal

Ship a private, invitation-only preview with one promise: keep a real task coherent while users move between Agents. Don't market workflow automation, universal compatibility, or public distribution readiness yet.
R1 · 05:38

HermesEvidence challenge

The beta should test behavior, not enthusiasm. Measure whether users resume the same task, bring in a second Agent for a clear reason, and use the trace before accepting a result. Without those signals, the category story is still unproven.
R2 · 05:44

OpenCodeReviewEngineering review

The MVP is stable enough for a controlled evaluation, but not for a public download. Tests and packaging pass; the blocker is trusted macOS signing and notarization. Keep the first release private and label unsupported platforms explicitly.
R2 · 05:52

YouYouReconciliation

Round 2: reconcile the disagreement. Keep only decisions we can verify this week and give me one acceptance gate.
Adoption

Human GateDecision

Ship V1.0.3 to the private preview only. No-Go for public release. Reopen after Developer ID signing and notarization pass on a clean Mac.

1 / 5

INSIDE THE WORK CELL

Inspectable from discovery to adoption.

Local Agent discovery

Pick the participants before any work runs. Meldwork detects installed CLIs and shows their adapter status.

Multi-agent review

Findings, evidence, and adoption decisions live in one Case. Dissent stays visible — never hidden behind a manager.

Auto Discussion V4

Propose, negotiate, verify — across rounds. Every Agent's position is preserved for later review.

Direct mode

One Agent, its native session preserved. Focused work with the full evidence trail still attached.

Local Agent discovery

Multi-agent review

Auto Discussion V4

Direct multimodal session with local files

Direct mode

Example: give Codex, Claude Code, and Gemini CLI the same task — review a PR, analyze a bug, or compare implementation approaches. Meldwork freezes the task context, sends it to all three agents, captures their independent findings as evidence, and lets you adopt only the result you approve. No manual context copying between terminals.

HOW MELDWORK DIFFERS

A General Agent workspace, not another terminal runner.

Meldwork is a local-first General Agent workspace for macOS. Coordinate research, analysis, writing, planning, review, and implementation across the Agent tools you already run, with no manual context copying between terminals.

Finding → Evidence → Decision → Disposition

Dissent stays visible. Disagreement, re-checks, and adoption records are tracked — never hidden behind a central manager.
Human Gate before workspace writes

Workspace writes are opt-in workflow controls. You adopt the result — the agents don't.
Local-first boundaries

Runs in your Electron work cell with existing CLIs. No cloud Agent fleet, no required server, conversations stay local.
Frozen context, independent judgment

Case-scoped independent judgments with evidence-backed decisions — not terminal count or worktree throughput.

WHERE MELDWORK FITS

How Meldwork compares to other Agent workspaces.

A condensed map of how Meldwork positions itself next to terminal runners, cloud fleets, communication networks, and framework-based orchestrators like CrewAI, LangGraph, and Claude Code Agent Teams. Full per-project comparison lives in the README ↗.

What Meldwork is, and isn't.

Is

A local-first Electron app for desktop review workflows.
A work cell that orchestrates the Agent CLIs you already have.
An evidence-first review record (Finding → Evidence → Decision → Disposition).
Workspace writes as opt-in workflow controls.
Source-available under the Meldwork Non-Commercial Source License 1.0.

Isn't

Not a hosted Agent fleet or general Agent framework.
Not automatic participant selection from a larger roster.
Not enterprise SSO / RBAC / governance tooling.
Not an operating-system sandbox — writes happen through your shell.
Not fully offline — a selected Agent may send prompts to its configured Provider.

ANSWERS

Questions teams ask before adopting.

Short, honest answers — the long versions live in architecture.md ↗.

Meldwork is a local-first General Agent workspace for macOS. It coordinates the Agent tools you already run from one workspace for research, analysis, writing, planning, review, and implementation. Meldwork freezes one task snapshot, captures independent findings as evidence, and puts a human adoption gate before any workspace change.
No, not by itself. Meldwork runs locally in your Electron work cell. Selected Agents send prompts, attachments, or Skills to their own configured Provider — but Meldwork itself is a local review UI, not a cloud relay. Conversations stay on your machine.
Twelve today — Codex, Claude Code, Gemini CLI, Qwen Code, Kimi Code, MiMo Code, OpenCode, OpenCodeReview, Hermes, Pi Agent, OpenClaw, and WorkBuddy. Approved Agent Connectors can be added through the Agent Connector SDK; custom executables use the desktop custom-Agent path.
Selected Agents independently propose, then critique with evidence, negotiate responsibilities, execute dependency-aware work, and verify the result. Every round is preserved — you can read the full Case later and adopt only the parts you approve.
A terminal runs one Agent at a time and shows raw output. tmux or zellij can show multiple agents side by side but require manual context copying between them. Meldwork scopes a single task across multiple agents, sends the same frozen snapshot to each, captures their independent findings as evidence, and keeps the human adoption decision visible as a first-class artifact. It also works across heterogeneous CLIs — you can mix Codex, Claude Code, Hermes, and OpenCode in one workspace, unlike Claude Code Agent Teams which only works within Claude Code.
Source-available under the Meldwork Non-Commercial Source License 1.0. You can inspect, run, modify, and share it for non-commercial purposes. Commercial use requires prior written permission — see COMMERCIAL_USE.md ↗.
The current V1.0.5 preview targets Apple silicon macOS (ad-hoc signed, not notarized — open Open Anyway in System Settings → Privacy & Security on first launch). Cross-platform builds are on the roadmap; the app is Electron so the same source compiles everywhere once those targets land.

General Agent Work, Legible & Accountable

Works with the agents you already run.

What is Meldwork?

Four steps from task to adopted result.

Select

Scope

Run

Review & adopt

Three ways agents work together.

Direct

Concurrent Responses

Auto Discussion V4

Four guarantees that hold the work together.

Human Gate

Frozen Context

Evidence Trail

Local-first Boundary

Launch Readiness Council — three rounds, one gate.

Inspectable from discovery to adoption.

Local Agent discovery

Multi-agent review

Auto Discussion V4

Direct mode

A General Agent workspace, not another terminal runner.

Finding → Evidence → Decision → Disposition

Human Gate before workspace writes

Local-first boundaries

Frozen context, independent judgment

How Meldwork compares to other Agent workspaces.

What Meldwork is, and isn't.

Is

Isn't

Questions teams ask before adopting.