Session Collab MCP
A provider-agnostic MCP server that synchronizes work-in-progress between multiple coding sessions to prevent file conflicts.
README
Session Collab MCP
A provider-agnostic Model Context Protocol (MCP) server that prevents conflicts when multiple agent sessions work on the same codebase simultaneously.
Problem
When using parallel coding-agent sessions or multi-agent workflows:
- Session A is refactoring some code
- Session B doesn't know and thinks the code "has issues" - deletes or reverts it
- Session A's work disappears
Root cause: No synchronization mechanism for "work intent" between sessions.
Solution
Session Collab MCP provides a Work-in-Progress (WIP) Registry that allows sessions to:
- Declare - Announce which files you're about to modify
- Check - Verify no other session is working on the same files
- Persist - Save context that survives context compaction
- Protect - Guard critical files from accidental changes
- Release - Free files when done
Positioning
session-collab-mcp has two layers:
- Core server: provider-agnostic MCP server over stdio or HTTP JSON-RPC
- Optional integrations: provider-specific packaging such as the Claude Code plugin in
plugin/
The core server should be the default mental model. Claude Code is one integration target, not the product boundary.
Installation
Option 1: Generic MCP Client over stdio
Use this with any MCP client that can launch a local stdio server:
{
"mcpServers": {
"session-collab": {
"command": "npx",
"args": ["-y", "session-collab-mcp@latest"]
}
}
}
The exact config wrapper depends on your MCP client, but the server contract is the same.
Option 2: HTTP Server + CLI
Use this when your client prefers MCP over HTTP JSON-RPC or when you want a generic shell-friendly wrapper:
# Start HTTP server
session-collab-http --host 127.0.0.1 --port 8765
# CLI wrapper (convenience REST client)
session-collab health
session-collab tools
session-collab doctor
session-collab call --name collab_session_start --args '{"project_root":"/repo","name":"demo"}'
For MCP-over-HTTP clients, use POST /mcp with JSON-RPC requests. The /v1/* endpoints are a convenience REST facade for lightweight automation and shell usage.
Option 3: Claude Code Plugin (Optional Integration)
Install as a Claude Code plugin only if Claude Code is your MCP client and you want automatic server setup, hooks, and skills:
# Add marketplace
/plugin marketplace add leaf76/session-collab-mcp
# Install plugin
/plugin install session-collab@session-collab-plugins
The plugin includes:
- MCP Server: Automatically configured
- Hooks: SessionStart, Stop, and PreCompact reminders
- Skills:
collab-startfor full initialization - Commands:
/session-collab:statusand/session-collab:end
Option 4: Global Installation
npm install -g session-collab-mcp
Features
Guided Session Workflow
The MCP tools give you a stable collaboration workflow across providers:
- Start a session with
collab_session_start - Check files with
collab_claim(action="check") - Reserve files with
collab_claim(action="create"); smart mode claims safe files and queues blocked files for coordination - Update visible progress with
collab_session_update - Save important context with
collab_memory_save - Release claims with
collab_claim(action="release") - End the session with
collab_session_end
Working Memory
Context persistence that survives context compaction:
- Findings: Bug root causes, investigation results
- Decisions: Architectural choices, design decisions
- State: Current implementation status
- Todos: Action items and tasks
- Important: Critical information to preserve
- Context: Background context for the session
File Protection
Guard important plan files or created files from accidental deletion:
- Register a protected plan with
collab_protect(action="register", type="plan", ...) - Register a created file with
collab_protect(action="register", type="file", ...) - Check protection status before deleting or replacing a file
Conflict Handling Modes
Configure behavior with collab_config:
| Mode | Behavior |
|---|---|
strict |
Block conflicting claims and require user coordination before editing |
smart (default) |
Claim non-conflicting files or symbols, queue blocked files, and expose coordination requests |
bypass |
Allow overlapping claims only with allow_conflicts=true, and return a warning |
In smart mode, prefer symbol-level claims when working inside a file that another session has claimed. Same-file work can proceed when claimed symbols do not overlap. If overlap cannot be proven safe, collab_claim(action="create") returns waiting_for_coordination or partial_claim_created with:
claimed_files: files reserved by the callersafe_files: files that can be edited nowblocked_files: files that need coordination firstcoordination_requests: queue entries visible throughcollab_statusandcollab_session_list
Auto-Release Options
| Option | Default | Description |
|---|---|---|
auto_release_immediate |
false |
Auto-release claims after Edit/Write |
auto_release_stale |
false |
Auto-release claims exceeding threshold |
stale_threshold_hours |
2 |
Hours before claim is considered stale |
auto_release_delay_minutes |
5 |
Grace period for stale release |
MCP Tools Reference
Session Management
| Tool | Purpose |
|---|---|
collab_session_start |
Register a new session |
collab_session_end |
End session and release all claims |
collab_session_list |
List active sessions with current task and active claim summaries |
collab_session_update |
Update heartbeat, current task, todos, and progress |
collab_config |
Configure session behavior |
collab_status |
Get session status summary |
Claims (1 unified tool)
| Tool | Actions |
|---|---|
collab_claim |
create, check, release, list (check: exclude_self defaults to true; create follows collab_config(mode)) |
Working Memory (3 tools)
| Tool | Purpose |
|---|---|
collab_memory_save |
Save context (upsert) |
collab_memory_recall |
Recall context |
collab_memory_clear |
Clear memories |
Protection (1 unified tool)
| Tool | Actions |
|---|---|
collab_protect |
register, check, list |
Status Monitoring
| Tool | Purpose |
|---|---|
collab_status |
Unified session status |
HTTP API (v1)
/v1/* endpoints map 1:1 to MCP tools and return JSON responses with trace_id on failures:
POST /v1/sessions/start→collab_session_startPOST /v1/sessions/end→collab_session_endPOST /v1/sessions/update→collab_session_updateGET /v1/sessions→collab_session_listPOST /v1/config→collab_configGET /v1/status→collab_statusPOST /v1/claims→collab_claim(create)POST /v1/claims/check→collab_claim(check)POST /v1/claims/release→collab_claim(release)GET /v1/claims→collab_claim(list)POST /v1/memory/save→collab_memory_savePOST /v1/memory/recall→collab_memory_recallPOST /v1/memory/clear→collab_memory_clearPOST /v1/protect/register→collab_protect(register)POST /v1/protect/check→collab_protect(check)GET /v1/protect/list→collab_protect(list)POST /v1/tools/call/GET /v1/tools(generic access)
MCP over HTTP
POST /mcpaccepts JSON-RPCinitialize,tools/list, andtools/callrequestsGET /mcpcurrently returns a clear "stream not supported" response instead of pretending to be full Streamable HTTP SSE- Localhost binds enforce Host and Origin validation
- Non-local binds require both
SESSION_COLLAB_HTTP_TOKENand an allowed-host list viaSESSION_COLLAB_ALLOWED_HOSTSor repeated--allowed-host
Usage Examples
Basic Workflow
# Session A starts working
collab_session_start(project_root="/my/project", name="feature-auth")
collab_claim(session_id="session-a", action="create", files=["src/auth.ts"], intent="Adding JWT support")
collab_session_update(session_id="session-a", current_task="Adding JWT support")
# Session B checks before editing
collab_claim(session_id="session-b", action="check", files=["src/auth.ts"])
# Result: "CONFLICT: src/auth.ts is claimed by 'feature-auth'"
# create in smart mode queues blocked files instead of editing over the owner
collab_claim(session_id="session-b", action="create", files=["src/auth.ts"], intent="Coordinated auth work")
# Result: status="waiting_for_coordination", blocked_files=["src/auth.ts"]
# If only a different symbol is needed, claim that symbol explicitly
collab_claim(
session_id="session-b",
action="create",
symbols=[{ file="src/auth.ts", symbols=["refreshToken"], symbol_type="function" }],
intent="Update refresh token"
)
# Result: status="created" if no claimed symbol overlaps
# If you want to include your own claims in the check
collab_claim(session_id="session-a", action="check", files=["src/auth.ts"], exclude_self=false)
# Session A finishes
collab_claim(session_id="session-a", action="release", claim_id="...")
Working Memory
# Save a finding
collab_memory_save(
session_id="abc123",
category="finding",
key="auth_bug_root_cause",
content="Missing token validation in refresh flow",
priority=80
)
# Recall active memories
collab_memory_recall(session_id="abc123", active=true)
File Protection
# Protect a plan document
collab_protect(
action="register",
session_id="abc123",
type="plan",
file_path="docs/feature-plan.md",
title="Feature plan",
content_summary="Steps, risks, and rollout notes"
)
# Check before editing
collab_protect(
action="check",
session_id="abc123",
file_path="docs/feature-plan.md"
)
# Result: "Protected (plan). Confirm before deleting."
Status Monitoring
# Get session status
collab_status(session_id="abc123")
# Result: {
# session: { id: "abc123", name: "feature-auth", status: "active" },
# claims: [...],
# other_sessions: 1,
# message: "Session active. 2 claim(s), 5 memories."
# }
Migration from v1.x
Version 2.0 introduces breaking changes with a simplified API. See MIGRATION.md for detailed migration instructions.
Key Changes
- Tool Consolidation: 50+ tools → 10 core tools
- Action-Based Interface: Single tools with multiple actions
- Simplified Responses: Cleaner, flatter response formats
- Removed Features: LSP integration, messaging, notifications, queuing
Data Storage
All data is stored locally in ~/.claude/session-collab/collab.db (SQLite).
- No remote server required
- Localhost HTTP usage works without an API token
- Non-local HTTP binds require
SESSION_COLLAB_HTTP_TOKEN - Works offline
- Uses WAL mode for multi-process safety
Development
Prerequisites
- Node.js 18+
- npm or yarn
Setup
npm install
npm run build
Legacy Build (Optional)
Legacy schemas/queries are kept out of the default bundle. To include a legacy entry for compatibility:
SESSION_COLLAB_INCLUDE_LEGACY=true npm run build
Maintenance note: legacy exports are for backward compatibility only and are not exposed in the v2 tool list.
Scripts
npm run build # Build with tsup
npm run start # Start the MCP server
npm run start:dev # Start in development mode
npm run typecheck # Run TypeScript type checking
npm run lint # Run ESLint
npm run test # Run tests with Vitest
npm run test:http # Run HTTP integration tests
npm run test:release # Run release gate: typecheck, lint, tests, HTTP tests, npm pack dry-run
HTTP Integration Tests
HTTP integration tests require a local listen port. Enable them with:
npm run test:http
HTTP CLI Doctor
When using the HTTP server, validate the running server with:
session-collab doctor --base-url http://127.0.0.1:8765
Historical Notes
The changelog entries below document historical milestones, including tools and workflows that were removed before the current v2 API. Treat the tool tables and examples above as the source of truth for the current public surface.
Project Structure
session-collab-mcp/
├── bin/ # Executable entry point
├── migrations/ # SQLite migration files
├── plugin/ # Optional Claude Code integration
├── src/
│ ├── cli.ts # CLI entry point
│ ├── constants.ts # Version and server instructions
│ ├── db/ # Database layer
│ ├── mcp/ # MCP protocol implementation
│ │ ├── tools/ # Tool implementations
│ │ │ ├── session.ts # Session management
│ │ │ ├── claim.ts # File/symbol claims
│ │ │ ├── memory.ts # Working memory
│ │ │ └── protection.ts # File protection
│ │ └── ...
│ └── utils/
└── package.json
Changelog
v2.3.1
- Add
collab_session_updatefor heartbeat, current task, todo, and progress reporting - Enrich
collab_session_listwith current task and active claim summaries - Make
collab_claim(action="create")respectcollab_config(mode)with smart coordination by default - Add
session-collab doctorfor HTTP server health and tool-surface checks - Add
npm run test:httpandnpm run test:releaserelease gates - Update dev test tooling to clear npm audit findings
v2.1.0
- Add HTTP server + CLI wrapper for universal AI CLI usage
- Add HTTP API endpoints and utils tests
- Add legacy entry for deprecated schemas/queries (optional build)
- Improve claim conflict accuracy and release summaries
- Expand test coverage across MCP tools and DB flows
v2.0.0 (Breaking)
- Major Simplification: Reduced from 50+ tools to 10 core tools
- Action-Based Design: Unified tools with action parameters
- Removed Features: LSP integration, messaging, notifications, queuing, decision tracking
- Improved Performance: Faster startup and reduced complexity
- Better Testing: Comprehensive test coverage for all tool actions
- Migration Guide: Detailed upgrade path from v1.x
v0.8.0
- Add working memory system for context persistence (
collab_memory_*tools) - Add plan protection (
collab_plan_register,collab_plan_update_status) - Add file protection (
collab_file_register,collab_file_check_protected) - Memory categories: finding, decision, state, todo, important, context
- Pinned memories survive context compaction
- Plan lifecycle: draft → approved → in_progress → completed → archived
v0.7.1
- Add
collab_auto_releasetool for releasing claims after editing - Add auto-release config options:
auto_release_immediate,auto_release_stale - Add
cleanupStaleClaims()for automatic stale claim cleanup - Add PostToolUse hook to remind auto-release after Edit/Write
v0.7.0
- Add priority system for claims (0-100 with levels: critical/high/normal/low)
- Add claim queue system (
collab_queue_join,collab_queue_leave,collab_queue_list) - Add notification system (
collab_notifications_list,collab_notifications_mark_read) - Add audit history tracking (
collab_history_list) - Add
collab_claim_update_priorityfor escalating urgent work
v0.6.0
- Optimize database queries with composite indexes
- Extract shared utilities (crypto, response builders)
- Remove unused auth and token modules
- Use precompiled JS for 15x faster startup
- Fix GROUP_CONCAT delimiter for multi-value queries
- Add unified Zod validation across tools
v0.5.0
- Add reference tracking and impact analysis (Phase 3)
- Add symbol-level claims and LSP integration
- Fix SQLite WAL sync for multi-process MCP servers
- Add
collab_configtool for conflict handling modes
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.