Imported from reason-machines/mcp-skills (skills/concord-mcp-workspace/SKILL.md). Install upstream with npx skills add reason-machines/mcp-skills --skill concord-mcp-workspace. Copyright stays with the author.

Concord MCP Workspace Skill

Skill by ara.so — MCP Skills collection.

Overview

Concord MCP provides a shared workspace for AI coding agents to coordinate work, preserve decisions, and hand off tasks cleanly. It's "Google Workspace for AI agents" — giving Claude Code, Codex, Cursor, and other MCP-capable agents one local-first place to track presence, ownership, task memory, and handoffs within a repository.

Core capabilities:

• Agent presence registration and roster tracking
• Task claiming with scoped ownership and version control
• Explicit handoffs with acceptance/decline workflows
• Task memory (decisions, assumptions, findings) attached to work
• Stale claim detection when agents disappear
• Review packet generation with scope, tests, risks, and provenance

Installation

Global CLI Installation

npm install -g @concord-ai/concord-mcp

Per-Client Setup

Cursor (one-click):

# Use the deeplink or:
concord install

Claude Code, Codex, other MCP clients:

concord install

The concord install command:

• Registers the MCP server in .mcp.json, .cursor/mcp.json, or ~/.codex/config.toml
• Writes agent instructions to CLAUDE.md, AGENTS.md, .codex/, .cursor/rules/
• Merges with existing config (safe to re-run)

Manual MCP registration (if using --no-mcp):

{
  "mcpServers": {
    "concord": {
      "command": "npx",
      "args": ["-y", "@concord-ai/concord-mcp@latest"]
    }
  }
}

Initialize Repository Workspace

cd /path/to/your/repo
concord init

Creates .concord/ directory with SQLite workspace (auto-added to .gitignore).

Environment Variables

# Optional: Set explicit repository root
export CONCORD_REPO_ROOT=/path/to/repo

# Optional: Restrict dynamic workspace joins (path-delimited)
export CONCORD_ALLOWED_ROOTS=/path/to/repo1:/path/to/repo2

# Optional: Disable CLI update checks
export CONCORD_NO_UPDATE_CHECK=1

MCP Tools Reference

Workspace & Presence

join_workspace

Join or select a repository workspace without restarting MCP.

{
  "repository_path": "/absolute/path/to/repo"
}
// Returns: { workspace_id: "ws_...", repository_root: "..." }

register_agent

Register agent presence and keep it alive.

{
  "agent_id": "cursor-agent-1",
  "agent_type": "cursor",
  "capabilities": ["code-generation", "review"]
  // Optional: workspace_id
}
// Returns: { agent_id, expires_at, ... }

get_work_state

Get roster, active tasks, overlaps, and stale claims.

{
  "agent_id": "cursor-agent-1"
  // Optional: workspace_id
}
// Returns: { agents: [...], tasks: [...], overlaps: [...], stale_claims: [...] }

Task Creation & Claiming

claim_work

Claim scoped work and create a task.

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "scope": {
    "files": ["src/auth.ts", "src/middleware/auth.ts"],
    "modules": ["authentication"],
    "description": "Refactor authentication to use JWT"
  }
  // Optional: workspace_id, metadata
}
// Returns: { task_id, version, status: "active", ... }

update_task

Record task memory (decisions, assumptions, findings).

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "context_update": {
    "decisions": ["Use jsonwebtoken library", "Store tokens in httpOnly cookies"],
    "assumptions": ["Token expiry set to 24 hours"],
    "findings": ["Existing auth uses sessions, needs migration"]
  }
  // Optional: workspace_id, expected_version
}

get_task_context

Read task memory and ownership history.

{
  "task_id": "refactor-auth"
  // Optional: workspace_id
}
// Returns: { context: {...}, ownership_history: [...] }

Task Assignment & Transfer

assign_task

Assign work to a named agent (requires acceptance).

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "assignee_id": "claude-code-agent",
  "expected_version": 1
  // Optional: workspace_id
}
// Task moves to "assigned" status, awaiting acceptance

accept_task

Accept an assigned task.

{
  "agent_id": "claude-code-agent",
  "task_id": "refactor-auth",
  "expected_version": 2
  // Optional: workspace_id
}
// Task moves to "active" with new owner

release_task

Release ownership without assigning.

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "expected_version": 1
  // Optional: workspace_id
}

reassign_task

Transfer to another agent with versioning.

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "new_assignee_id": "codex-agent",
  "expected_version": 1
  // Optional: workspace_id
}

Handoffs

offer_handoff

Deliver handoff to named recipient with acceptance workflow.

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "recipient_id": "claude-code-agent",
  "handoff_data": {
    "summary": "JWT auth refactor complete",
    "completed_work": ["Implemented JWT signing/verification", "Added middleware"],
    "remaining_work": ["Write integration tests", "Update documentation"],
    "files_changed": ["src/auth.ts", "src/middleware/auth.ts"],
    "tests_added": ["src/auth.test.ts"],
    "risks": ["Session migration not handled"]
  },
  "expected_version": 1
  // Optional: workspace_id, expires_in_seconds (default 300)
}

accept_handoff

Accept offered handoff.

{
  "agent_id": "claude-code-agent",
  "task_id": "refactor-auth"
  // Optional: workspace_id
}

decline_handoff

Decline offered handoff.

{
  "agent_id": "claude-code-agent",
  "task_id": "refactor-auth",
  "reason": "Need to finish current task first"
  // Optional: workspace_id
}

handoff

Record completion/review evidence without transferring ownership.

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "handoff_data": {
    "summary": "Ready for review",
    "completed_work": ["All implementation done"],
    "files_changed": ["src/auth.ts"],
    "tests_added": ["src/auth.test.ts"]
  }
  // Optional: workspace_id
}

Task Closure

close_task

Mark task complete or closed.

{
  "agent_id": "cursor-agent-1",
  "task_id": "refactor-auth",
  "outcome": "complete",
  "expected_version": 3
  // Optional: workspace_id
}

reopen_task

Reopen a closed task (audited).

{
  "agent_id": "claude-code-agent",
  "task_id": "refactor-auth",
  "reason": "Found edge case in production"
  // Optional: workspace_id
}

CLI Commands

Workspace Management

# Initialize workspace in current repo
concord init

# Select workspace by repository path
concord --repo /path/to/project status

# Select workspace by ID
concord --workspace ws_abc123 status

Status & Monitoring

# View roster, active work, overlaps, stale claims
concord status

# Live dashboard (TUI)
concord dashboard
# Tab: change panes | j/k: navigate | /: filter | ?: help | q: quit

# List active agents and their work
concord who

# List all tasks
concord tasks

Handoff & Review

# Print latest handoff for task
concord handoff refactor-auth

# Print review packet
concord review-packet refactor-auth

# Export markdown artifacts to .concord/
concord export markdown

Diagnostics

# Run workspace health checks
concord doctor

Common Patterns

Multi-Agent Workflow

// Agent 1: Claim and start work
await use_mcp_tool("concord", "register_agent", {
  agent_id: "agent-1",
  agent_type: "cursor",
  capabilities: ["code-generation"]
});

await use_mcp_tool("concord", "claim_work", {
  agent_id: "agent-1",
  task_id: "add-logging",
  scope: {
    files: ["src/logger.ts"],
    description: "Add structured logging"
  }
});

// Record decisions as you work
await use_mcp_tool("concord", "update_task", {
  agent_id: "agent-1",
  task_id: "add-logging",
  context_update: {
    decisions: ["Use winston library", "Log to both console and file"],
    assumptions: ["Log rotation handled by external service"]
  }
});

// Agent 2: Register and check for overlaps
await use_mcp_tool("concord", "register_agent", {
  agent_id: "agent-2",
  agent_type: "claude-code"
});

const state = await use_mcp_tool("concord", "get_work_state", {
  agent_id: "agent-2"
});
// Check state.overlaps before claiming similar scope

// Agent 1: Offer handoff when ready
await use_mcp_tool("concord", "offer_handoff", {
  agent_id: "agent-1",
  task_id: "add-logging",
  recipient_id: "agent-2",
  handoff_data: {
    summary: "Logging implementation complete, needs tests",
    completed_work: ["Implemented winston logger", "Added config"],
    remaining_work: ["Write unit tests", "Add error scenarios"],
    files_changed: ["src/logger.ts", "src/config/logger.ts"]
  },
  expected_version: 1
});

// Agent 2: Accept handoff
await use_mcp_tool("concord", "accept_handoff", {
  agent_id: "agent-2",
  task_id: "add-logging"
});

Task Context Preservation

// Continuously update context as you learn
await use_mcp_tool("concord", "update_task", {
  agent_id: "my-agent",
  task_id: "optimize-queries",
  context_update: {
    findings: [
      "Query uses N+1 pattern in user.posts",
      "Database missing index on posts.user_id"
    ],
    decisions: [
      "Add eager loading with includes",
      "Create migration for index"
    ]
  }
});

// Later: Read context before continuing
const context = await use_mcp_tool("concord", "get_task_context", {
  task_id: "optimize-queries"
});
// Use context.decisions and context.findings to continue work

Review-Ready Handoff

// Complete work and prepare review packet
await use_mcp_tool("concord", "handoff", {
  agent_id: "my-agent",
  task_id: "api-versioning",
  handoff_data: {
    summary: "Added v2 API with backwards compatibility",
    completed_work: [
      "Implemented /api/v2 routes",
      "Added version middleware",
      "Maintained v1 compatibility"
    ],
    files_changed: [
      "src/routes/v2/*.ts",
      "src/middleware/version.ts"
    ],
    tests_added: [
      "tests/api/v2/users.test.ts",
      "tests/middleware/version.test.ts"
    ],
    risks: [
      "V1 deprecation timeline not finalized",
      "Migration guide needed for clients"
    ],
    notes: "Consider adding deprecation warnings to v1 responses"
  }
});

// Close when reviewed and merged
await use_mcp_tool("concord", "close_task", {
  agent_id: "my-agent",
  task_id: "api-versioning",
  outcome: "complete",
  expected_version: 2
});

Handling Stale Claims

# Check for stale claims
concord status

# From CLI, manually release if needed:
# (MCP agents should monitor get_work_state for stale_claims)

// In agent: Detect and handle stale claims
const state = await use_mcp_tool("concord", "get_work_state", {
  agent_id: "my-agent"
});

if (state.stale_claims.length > 0) {
  // Option 1: Claim abandoned work
  const stale = state.stale_claims[0];
  await use_mcp_tool("concord", "claim_work", {
    agent_id: "my-agent",
    task_id: stale.task_id,
    scope: stale.scope
  });
  
  // Option 2: Read context and continue
  const context = await use_mcp_tool("concord", "get_task_context", {
    task_id: stale.task_id
  });
}

Workspace Structure

.concord/
├── concord.db          # SQLite source of truth
├── HANDOFF.md          # Human-readable latest handoff
├── REVIEW_PACKET.md    # Generated review packet
└── WORK_STATE.json     # Optional exported state

Troubleshooting

MCP Server Not Recognized

Symptom: Agent can't see Concord tools

Solution:

# Re-run install
concord install

# Verify registration
cat .cursor/mcp.json  # or .mcp.json

# Restart your AI agent/IDE

Workspace Not Found

Symptom: "No workspace selected" errors

Solution:

# Ensure you're in a repo with .concord/
ls .concord/

# Or initialize
concord init

# Or set explicit root
export CONCORD_REPO_ROOT=$(pwd)

# Or use join_workspace tool
await use_mcp_tool("concord", "join_workspace", {
  repository_path: "/absolute/path/to/repo"
});

Version Conflicts

Symptom: "Expected version X but found Y" errors

Solution:

// Always read current version first
const context = await use_mcp_tool("concord", "get_task_context", {
  task_id: "my-task"
});

// Use the current version
await use_mcp_tool("concord", "close_task", {
  agent_id: "my-agent",
  task_id: "my-task",
  expected_version: context.version,
  outcome: "complete"
});

Agent Presence Expired

Symptom: Agent shows as offline in roster

Solution:

// Register more frequently (any write operation extends presence)
await use_mcp_tool("concord", "register_agent", {
  agent_id: "my-agent",
  agent_type: "cursor"
});

// Or make presence implicit by working
await use_mcp_tool("concord", "update_task", {
  agent_id: "my-agent",  // Keeps presence alive
  task_id: "my-task",
  context_update: { /* ... */ }
});

Linked Worktree Issues

Symptom: Different .concord/ in worktree vs main checkout

Solution: Concord automatically follows Git's commondir to share workspace. If not working:

# Verify Git worktree setup
git worktree list

# Ensure .git/commondir exists in worktree
cat .git/commondir

# Both should resolve to same .concord/ location

Upgrade Issues

# Update global package
npm install -g @concord-ai/concord-mcp@latest

# Verify version
concord --version

# Migrations run automatically on next workspace access

Best Practices

1. Register early: Call register_agent when starting work
2. Check overlaps: Use get_work_state before claiming new scope
3. Update context frequently: Record decisions/findings as you go
4. Use versions: Always pass expected_version for lifecycle changes
5. Explicit handoffs: Use offer_handoff + accept_handoff for clean transfers
6. Review packets: Call handoff with comprehensive data before marking complete
7. Monitor stale claims: Check get_work_state.stale_claims regularly