Your AI agents forget everything. Twining remembers.
Persistent project memory for Claude Code and other MCP clients.

You spend two hours with Claude Code making architectural decisions. You choose PostgreSQL over MongoDB. You settle on JWT for auth. You flag a race condition in the payment module. Then the session ends.

Tomorrow you start a new session. Claude has no idea what happened. The decisions are gone. The warnings are gone. The rationale is gone. You re-explain everything — or worse, Claude silently contradicts yesterday's choices.

This gets worse with multiple agents. Agent A decides on REST. Agent B picks gRPC for the same service. Neither knows the other exists. You find out when the code doesn't compile.

Context windows are ephemeral. Your project's decisions shouldn't be.

Twining is an MCP server that gives your AI agents persistent project memory. Decisions survive context resets. New sessions start informed. Multi-agent work stays coordinated.

# Install in 10 seconds
/plugin marketplace add daveangulo/twining-mcp
/plugin install twining@twining-marketplace

Record a decision with rationale:

> Use twining_decide to record: chose PostgreSQL over MongoDB for ACID compliance

Twining captures the decision, rationale, alternatives considered, confidence level, and affected files — as a structured record, not chat history.

Start a new session. Get caught up instantly:

> Use twining_assemble for the database module

Twining scores every decision, warning, and finding by relevance to your task, then fills a token budget in priority order. You get exactly the context you need — no firehose, no re-explaining.

Ask why things are the way they are:

> Use twining_why src/auth/middleware.ts

Returns the full decision chain for any file: what was decided, when, why, what alternatives were rejected, and which commit implemented it.

CLAUDE.md is static. You write it once and update it manually. It doesn't capture decisions as they happen, doesn't track rationale or alternatives, doesn't detect conflicts between agents, and can't selectively assemble context within a token budget.

Twining is dynamic. Every twining_decide call records a structured decision. Every twining_post shares a finding or warning. Every twining_assemble scores relevance and delivers precisely what the current task needs. The .twining/ directory is your project's living institutional memory.

Orchestrators (like agent swarms and hierarchical coordinators) route work by assigning tasks. Twining coordinates by sharing state. The difference matters:

• Orchestrators hold coordination context in their own context window — a single point of failure that degrades as the window fills
• Twining's blackboard persists coordination state outside any agent's window, surviving context resets without information loss

Agents self-select into work by reading the blackboard. No central bottleneck. No relay that drops context. Every agent sees every other agent's decisions and warnings, directly.

# Add the marketplace (one-time)
/plugin marketplace add daveangulo/twining-mcp

# Install the plugin
/plugin install twining@twining-marketplace

Includes the MCP server, 8 skills, lifecycle hooks, coordinator subagent, and commands. Skills teach Claude when and how to use Twining — everything works automatically.

Commit this to your repo's .claude/settings.json so every team member gets Twining on clone:

{
  "extraKnownMarketplaces": {
    "twining-marketplace": {
      "source": {
        "source": "github",
        "repo": "daveangulo/twining-mcp"
      }
    }
  },
  "enabledPlugins": {
    "twining@twining-marketplace": true
  }
}

When team members trust the repository folder, Claude Code automatically installs the marketplace and plugin.

For non-Claude-Code clients (Cursor, Windsurf, etc.):

claude mcp add twining -- npx -y twining-mcp --project .

Or add to .mcp.json:

{
  "mcpServers": {
    "twining": {
      "command": "npx",
      "args": ["-y", "twining-mcp", "--project", "."]
    }
  }
}

MCP server instructions are included automatically in the initialize response.

If you previously configured Twining manually, switch to the plugin:

1. Remove manual MCP server: claude mcp remove twining
2. Install plugin: /plugin marketplace add daveangulo/twining-mcp then /plugin install twining@twining-marketplace
3. Clean up: remove Twining hooks from .claude/settings.json, remove .claude/agents/twining-aware.md if present, remove Twining sections from CLAUDE.md (skills handle this now)
4. Keep: .twining/ directory (all state preserved)
5. Verify: /twining:status

The plugin handles agent instructions automatically via skills. For the MCP-only install path, add Twining instructions to your project's CLAUDE.md so agents use it automatically — see docs/CLAUDE_TEMPLATE.md for a ready-to-copy template.

A web dashboard starts automatically at http://localhost:24282 — browse decisions, blackboard entries, knowledge graph, and agent state. Configurable via TWINING_DASHBOARD_PORT.

Stats overview: blackboard entries, decisions, graph entities, and activity breakdown

Interactive knowledge graph: files, decisions, classes, and their relationships

Decisions auto-populate the graph: twining_decide creates file and function entities with decided_by relations for every affected file.

All state lives in .twining/ as plain files — JSONL for the blackboard, JSON for decisions, graph, agents, and handoffs. Everything is jq-queryable, grep-able, and git-diffable. No database. No cloud. No accounts.

Architecture layers:

• Storage — File-backed stores with locking for concurrent access
• Engine — Decision tracking, blackboard, graph traversal, context assembly with token budgeting, agent coordination
• Embeddings — Local all-MiniLM-L6-v2 via @huggingface/transformers, lazy-loaded, with keyword fallback. The server never fails to start because of embedding issues.
• Dashboard — Read-only web UI with cytoscape.js graph visualization and vis-timeline
• Tools — MCP tool definitions validated with Zod, mapping 1:1 to the tool surface

See TWINING-DESIGN-SPEC.md for the full specification.

Does Twining slow down Claude Code? No. It's a local MCP server — tool calls are local file reads/writes. Semantic search loads lazily on first use.

Can I use it with Cursor, Windsurf, or other MCP clients? Yes. Twining is a standard MCP server. Any MCP host can connect to it.

Where does my data go? All coordination state is local in .twining/. Tool call metrics are stored locally in .twining/metrics.jsonl (gitignored). Optional anonymous telemetry can be enabled — see Analytics below.

Is Twining an agent orchestrator? No. It's a coordination state layer. It captures what agents decided and why, and makes that knowledge available to future agents. Use it alongside orchestrators, agent teams, or standalone sessions.

Twining includes a three-layer analytics system to help you understand the value it provides.

The web dashboard includes an Insights tab showing:

• Value Metrics — Blind decision prevention rate, warning acknowledgment, test coverage via tested_by graph relations, commit traceability, decision lifecycle, knowledge graph stats, and agent coordination metrics
• Tool Usage — Call counts, error rates, average/P95 latency per tool
• Error Breakdown — Errors grouped by tool and error code

All value metrics are computed from existing .twining/ data — no new data collection needed.

Every MCP tool call is automatically instrumented with timing and success/error tracking. Metrics are stored locally in .twining/metrics.jsonl (gitignored — operational data, not architectural).

To disable local metrics collection, set in .twining/config.yml:

analytics:
  metrics:
    enabled: false

Anonymous aggregate usage data can optionally be sent to PostHog to help improve Twining. Disabled by default. To enable, add to .twining/config.yml:

analytics:
  telemetry:
    enabled: true

That's it — the PostHog project key is built into the source code. If you run your own PostHog instance, you can override with posthog_api_key and posthog_host.

What is sent: tool names, call durations, success/failure booleans, server version, OS, architecture.

What is never sent: file paths, decision content, agent names, error messages, tool arguments, environment variables.

Privacy safeguards:

• DO_NOT_TRACK=1 environment variable always overrides config
• CI=true auto-disables telemetry
• Identity is a SHA-256 hash of hostname + project root (never raw paths)
• Network failures are silent — no retries
• posthog-node is an optional dependency — graceful no-op if not installed

npm install       # Install dependencies
npm run build     # Build
npm test          # Run tests (600+ tests)
npm run test:watch

Requires Node.js >= 18.

Two GitHub Actions workflows automate build verification and publishing:

CI (.github/workflows/ci.yml) — runs on every PR and push to main:

• Builds and tests across Node 18, 20, and 22
• Cancels in-progress runs when a new push arrives on the same branch

Publish (.github/workflows/publish.yml) — runs on v* tag push:

• Builds with POSTHOG_API_KEY baked in for published packages
• Runs the full test suite as defense-in-depth
• Publishes to npm with --provenance for supply chain attestation
• Creates a GitHub Release with auto-generated release notes
• Supports manual trigger via workflow_dispatch with a dry-run option

To publish a new version:

npm version patch   # or minor, major
git push && git push --tags

Required secrets (configured in GitHub repo Settings > Secrets):

MIT