Wormhole

Keep your AI coding agents in sync. Wormhole gives Claude Code, GitHub Copilot, and Cursor a shared memory layer—so when you switch tools mid-task, nothing gets lost.

- 🔀Multiple subagentswithin the same tool (e.g., parallel Claude tasks)
- 🔄Different AI toolsentirely (Claude ↔ Copilot ↔ Cursor)

⚠️Disclaimer: Wormhole is an early-stage project. APIs and behavior may change, and there may be rough edges. It’s built in the open and evolving fast based on real developer feedback.

- Universal Logging- Singlelogtool for any action type
- Event Tagging- Categorize events with tags for better organization
- Session Management- Named work sessions with isolation
- Token Optimized- Compact output, delta queries, relevance filtering
- Conflict Detection- Know when agents touch the same files
- Stale Event Rejection- Automatically filters out file edits that no longer exist in the current project state
- Web UI Visualization- View sessions, timeline events, and insights withnpx wormhole ui
- Knowledge Capture & Search- Save decisions/pitfalls and surface them with intent-aware search

Try instantly with npx (no installation required):
- start_session
- Pull context:search_project_knowledge+get_recent
- Before edits:check_conflicts
- During work:logevery file_edit/cmd_run/decision/test_result/todos
- Capture learnings:save_knowledge(decision/pitfall/convention/constraint)
- Finish:end_sessionwith summary

start_session({ project_path: ".", agent_id: "copilot", name: "fix-auth" }) search_project_knowledge({ project_path: ".", intent: "debugging", query: "auth" }) get_recent({ project_path: "." }) check_conflicts({ project_path: ".", files: ["src/auth.ts"] }) log({ action: "file_edit", agent_id: "copilot", project_path: ".", content: { file_path: "src/auth.ts", description: "Fix timeout" } }) save_knowledge({ project_path: ".", knowledge_type: "decision", title: "Use async DB client", content: "Prevents blocking" }) end_session({ session_id: "abc-123", summary: "Auth fixed; tests green" })

Visualize your agent activity with the built-in web interface:

# Start the UI server (default port: 3000) npx wormhole ui # Or specify a custom port npx wormhole ui 8080

Then openhttp://localhost:3000in your browser to see:

- 📊Dashboard- Stats on events, sessions, and agents
- ⏱️Timeline- Visual event stream with filtering
- 📋Sessions- All work sessions with details
- 📈Insights- Action types and tag analytics

Claude Code— Add to~/.claude/claude_code_config.json:

{ "mcpServers": { "wormhole": { "command": "npx", "args": ["-y", "wormhole-mcp"] } } }

GitHub Copilot— Add to.vscode/mcp.jsonin your project:

{ "servers": { "wormhole": { "command": "npx", "args": ["-y", "wormhole-mcp"] } } }

Then use"command": "wormhole-mcp"in your config.

git clone https://github.com/fatmali/wormhole.git cd wormhole npm install npm run build

Use"command": "node"with"args": ["/path/to/wormhole/dist/server.js"].

### Claude Code Plugin For Claude Code users, there's an optional plugin that bundles the MCP server config with a skill: 

bash # Install the plugin claude /install-plugin ./node_modules/wormhole-mcp/plugins/wormhole

claude --plugin-dir ./node_modules/wormhole-mcp/plugins/wormhole

Then invoke with/wormhole:wormholein Claude Code.

cp -r node_modules/wormhole-mcp/skills/wormhole .claude/skills/

- cmd_run- Command executions

- file_edit- File modifications

- decision- Design decisions with rationale

- test_result- Test outcomes

- feedback- User acceptance/rejection

- todos- Task items with status tracking

- plan_output- Planning artifacts (design, architecture, tasks)

- Any custom type you need

Get recent activity (compact by default):

get_recent({ project_path: "/path/to/project" })

[5m] claude: npm test → ✓ [8m] cursor: edit auth.ts "Add JWT" [12m] copilot: decided "Use Zod for validation" [15m] claude: auth.test.ts ✓ cursor: evt_123

- limit- Max events (default: 5)

- detail-minimal|normal|full

- since_cursor- Only new events (delta query)

- related_to- Filter by file paths

- action_types- Filter by action types

- tags- Filter by tags (e.g.,["bugfix", "feature"])

Get all unique tags used in a project with counts:

get_tags({ project_path: "/path/to/project" }) // Output: // tags: // bugfix (12) ### save_knowledge Persist decisions, pitfalls, conventions, or constraints so agents don’t repeat mistakes. `javascript save_knowledge({ project_path: "/path/to/project", knowledge_type: "pitfall", title: "Avoid fs.readFileSync in handlers", content: "Blocks event loop; causes timeouts", confidence: 0.9 })

Intent-aware lookup of stored knowledge. Prefers types that match your intent.

search_project_knowledge({ project_path: "/path/to/project", intent: "debugging", query: "auth" }) // → [{ type: "pitfall", summary: "Avoid fs.readFileSync", confidence: 0.9 }]

// feature (8) // testing (5) // auth (3)

Options: - with_counts - Include event counts per tag (default: true) ### check_conflicts Detect concurrent file edits: `javascript check_conflicts({ project_path: "/path/to/project" })

Wormhole automatically tracks and validates file edits to ensure agents never act on stale information. When afile_editevent is logged with adiff, Wormhole:

- Extracts the full patch- Stores all added/removed lines from the diff

- Validates on query- When events are retrieved viaget_recentor conflict detection, each file edit is checked against the current file state

- Fuzzy matching- Uses intelligent matching to handle code that moved positions, only rejecting truly stale edits

- Auto-filters- Rejected events are automatically excluded from results

log({ action: "file_edit", agent_id: "claude-code", project_path: "/path/to/project", content: { file_path: "src/auth.ts", description: "Added JWT validation", diff: --- a/src/auth.ts +++ b/src/auth.ts @@ -10,6 +10,7 @@ function validateToken(token: string) { + const decoded = jwt.verify(token, SECRET); return decoded; } } })

Wormhole stores the full diff in the payload. Later, when another agent queries recent events:

- File still has the change→ Event is included

- Code was removed or changed→ Event is silently filtered out

- Code moved to different location→ Still recognized (fuzzy match)
Note:Thedifffield is NOT truncated (unlike other content fields), ensuring accurate validation even for large changes.
This ensures agents always work with accurate context about what's currently in the codebase.
The patch validation uses intelligent fuzzy matching:
- Checks if the added code exists anywhere in the current file

- Uses normalized comparison (trimmed whitespace)

- Accepts partial matches (code that contains or is contained by the search)

- Requires 60% of added lines to match for validation
- If a "removed" line still exists in the file → patch is stale

- This catches cases where a deletion was reverted
- File deleted: Patch fails validation

- Code refactored: Fuzzy matching still finds the logic if it exists

- Whitespace changes: Normalized comparison ignores formatting

- Line movements: Searches entire file, not just original position

- No patch stored: Event is kept (backward compatibility)

- Already rejected: Event is skipped on subsequent queries
- Validation runs only when events are queried (lazy evaluation)

- File I/O is cached by OS for repeated reads

- Minimal overhead: ~1-5ms per file_edit event

- Database stores full diffs efficiently as TEXT columns

// Clean entire project cleanup({ scope: "project", project_path: "/path/to/project" }) // Clean specific session cleanup({ scope: "session", session_id: "abc-123" }) // Clean everything cleanup({ scope: "all", force: true })

start_session({ project_path: "/path/to/project", agent_id: "claude-code", name: "bugfix-auth", description: "Fixing login timeout issue" }) // → session started: bugfix-auth (abc-123-def)

Sessions automatically isolate context—previous events hidden from queries.

end_session({ session_id: "abc-123-def", summary: "Fixed timeout by optimizing DB query" })

list_sessions({ project_path: "/path/to/project" })

● bugfix-auth (2h) by claude ○ feature-payment (1d) by cursor

switch_session({ session_id: "xyz-789" })

{ "retention_hours": 24, "max_payload_chars": 200, "auto_cleanup": true, "default_detail": "minimal", "default_limit": 5 }

Note:Themax_payload_charssetting truncates most content fields for display, butdifffields infile_editevents are always stored in full to enable accurate stale-event validation.

Wormhole minimizes token usage forget_recentresponses through four key strategies:
Instead of returning raw JSON, events are formatted as single-line summaries:

# Compact (default) - ~35 chars per event [5m] claude: npm test → ✓ # vs Full JSON - ~200+ chars per event {"id":42,"agent_id":"claude-code","action":"cmd_run","payload":"{\"command\":\"npm test\",\"exit_code\":0}","timestamp":1706621234567,"project_path":"/path/to/project","session_id":"abc-123"}

Thedetailparameter controls verbosity:

- minimal(default) — Single-line summaries with symbols (✓/✗)

- normal— Multi-line with key details

- full— Complete JSON payloads
Content fields are truncated to 200 characters by default (max_payload_charsconfig):

// Stored/displayed as: "Added authentication middleware with JWT validation and refresh token..." // Instead of full 2000+ char description

Exception:difffields infile_editevents are never truncated—they're needed for stale event validation.

Usesince_cursorto fetch only events since your last query:

// First call returns events + cursor get_recent({ project_path: "." }) // → [5 events] + cursor: evt_42 // Subsequent call returns only NEW events get_recent({ project_path: ".", since_cursor: "evt_42" }) // → [0-2 events] instead of repeating all 5

This prevents re-sending the same context repeatedly.

- default_limit: 5— Returns only 5 most recent events

- Agents can increase withlimitparam when needed

{ "max_payload_chars": 200, "default_detail": "minimal", "default_limit": 5 }

┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ Claude Code │ │ Copilot │ │ Cursor │ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │ │ │ └────────────────┼────────────────┘ │ ┌───────▼───────┐ │ Wormhole │ │ MCP Server │ └───────┬───────┘ │ ┌───────▼───────┐ │ SQLite │ │ timeline.db │ └───────────────┘
``

- Database:~/.wormhole/timeline.db
- Config:~/.wormhole/config.json
- Archives:~/.wormhole/archives/`

This is a web browser that enables your coding agent, such as Claude Code, to visit websites on your behalf and assist you in identifying bugs or creating UI test cases.

Coordination layer for AI agents across isolated developer local environments (devices)

⚡ Boost Requirement Analysis Efficiency by 200%! The World's First Team Collaboration MCP Server Designed for the AI Coding Era. Automatically analyzes requirements, generates full-stack code, and downloads design assets.

Multi-agent coordination over MCP: atomic gap-free claims, file leases, a shared ledger, presence, handoffs, and a task graph. Remote Streamable HTTP; self-host (AGPL) or hosted.

Inline review comments for markdown specs and design docs. Agents request human review mid-task via MCP and pause until you send feedback.

A pyRevit-based MCP server for Autodesk Revit, enabling connection to any MCP-compatible client.

AI-first file sharing and collaboration. 251 MCP tools give agents a full workspace: file storage, branded shares, comments, workflows, and built-in RAG. 50GB free, no credit card.

A file mover tool that stages and executes file moves safely. Works as both a CLI tool and an MCP server for AI agents.

The project tracker for teams of humans and AI agents — in plain files. MCP server + CLI, zero dependencies.

One shared context layer for AI agents and humans — live API specs, DB schemas, and versioned contracts across repos so every agent and teammate works from the same source of truth.