Thoughtbox

Multi-agent collaborative reasoning that's auditable. Thoughtbox is an MCP server for structured, multi-agent reasoning, with a companion web app for workspace and inspection flows. Every step is recorded as a structured thought in a persistent reasoning ledger that can be visualized, exported, and analyzed.

Runtime modes: Local development can use filesystem or in-memory storage. Deployed mode uses Supabase-backed storage, and the current production MCP server runs on Cloud Run.

Observatory UI showing a reasoning session with 14 thoughts and a branch exploration (purple nodes 13-14) forking from thought 5.

Code Mode

Thoughtbox exposes exactly two MCP tools using the Code Mode pattern:

• thoughtbox_search — Write JavaScript to query the operation/prompt/resource catalog. The LLM has full programmatic filtering power over the catalog.
• thoughtbox_execute — Write JavaScript using the tb SDK to chain operations. Access thoughts, sessions, knowledge, notebooks, hub, observability, and protocol tools through a unified namespace.

Workflow: search to discover available operations, then execute code against them. Use console.log() for debugging — output is captured in response logs.

This replaces per-operation tool registration with a two-tool surface that scales without context window bloat.

Multi-Agent Collaboration

The Hub is the coordination layer. Agents register with role-specific profiles, join shared workspaces, and work through a structured problem-solving workflow — all via thoughtbox_execute.

The workflow: register → create workspace → create problem → claim → work → propose solution → peer review → merge → consensus

Workspace primitives:

• Problem — A unit of work with dependencies, sub-problems, and status tracking (open → in-progress → resolved → closed)
• Proposal — A proposed solution with a source branch reference and review workflow
• Consensus — A decision marker tied to a thought reference for traceability
• Channel — A message stream scoped to a problem for discussion

Agent Profiles: MANAGER, ARCHITECT, DEBUGGER, SECURITY, RESEARCHER, REVIEWER — each provides domain-specific mental models and behavioral priming.

28 operations across identity, workspace management, problems, proposals, consensus, channels, and status reporting.

Auditable Reasoning

Every thought is a node in a graph — numbered, timestamped, linked to its predecessors, and persisted across sessions. This creates an auditable trail of how conclusions were reached.

Agents can think forward, plan backward, branch into parallel explorations, and revise earlier conclusions. Each pattern is a first-class operation:

Each thought carries a semantic thoughtType (reasoning, decision_frame, action_report, belief_snapshot, assumption_update, context_snapshot, progress) that classifies what kind of thought it is, orthogonal to the process pattern used.

See the Patterns Cookbook for comprehensive examples.

Real-Time Observability

The Observatory is a built-in web UI at http://localhost:1729 for watching reasoning unfold live.

• Live Graph — thoughts appear as nodes in real-time via WebSocket
• Branch Navigation — branches collapse into clickable stubs; drill in and back out
• Detail Panel — click any node to view full thought content
• Multi-Session — switch between active reasoning sessions
• Deep Analysis — analyze sessions for reasoning patterns, cognitive load, and decision points

The full observability stack includes OpenTelemetry tracing, Prometheus metrics, and Grafana dashboards.

Knowledge & Reasoning Tools

Knowledge Graph — Persistent memory across sessions. Capture insights, concepts, workflows, and decisions as typed entities with typed relations (BUILDS_ON, CONTRADICTS, SUPERSEDES, etc.) and visibility controls (public, agent-private, team-private).

Notebooks — Interactive literate programming combining documentation with executable JavaScript/TypeScript in isolated environments.

Client Compatibility

Thoughtbox is currently optimized for Claude Code. We are actively working on supporting additional MCP clients. Due to variation in capability support across the MCP ecosystem — server features (prompts, resources, tools), client features (roots, sampling, elicitation), and behaviors like listChanged notifications — we implement custom adaptations for many clients.

If you're using a client other than Claude Code and encounter issues, please open an issue describing your client and the problem.

Installation

Thoughtbox runs as a Docker-based MCP server. It requires Docker and Docker Compose.

Quick Start

git clone https://github.com/Kastalien-Research/thoughtbox.git
cd thoughtbox
docker compose up --build

This starts Thoughtbox and the full observability stack. The MCP server listens on port 1731 and the Observatory UI is available at http://localhost:1729.

MCP Client Configuration

Since Thoughtbox uses HTTP transport, configure your MCP client to connect via URL.

Claude Code

Add to your ~/.claude/settings.json or project .claude/settings.json:

{
  "mcpServers": {
    "thoughtbox": {
      "url": "http://localhost:1731/mcp"
    }
  }
}

To connect through the observability sidecar (adds OpenTelemetry tracing):

{
  "mcpServers": {
    "thoughtbox": {
      "url": "http://localhost:4000/mcp"
    }
  }
}

Cline / VS Code

Add to your MCP settings or .vscode/mcp.json:

{
  "servers": {
    "thoughtbox": {
      "url": "http://localhost:1731/mcp"
    }
  }
}

Usage Examples

Forward Thinking — Problem Analysis

Thought 1: "Users report slow checkout. Let's analyze..."
Thought 2: "Data shows 45s average, target is 10s..."
Thought 3: "Root causes: 3 API calls, no caching..."
Thought 4: "Options: Redis cache, query optimization, parallel calls..."
Thought 5: "Recommendation: Implement Redis cache for product data"

Backward Thinking — System Design

Thought 8: [GOAL] "System handles 10k req/s with <100ms latency"
Thought 7: "Before that: monitoring and alerting operational"
Thought 6: "Before that: resilience patterns implemented"
Thought 5: "Before that: caching layer with invalidation"
...
Thought 1: [START] "Current state: 1k req/s, 500ms latency"

Branching — Comparing Alternatives

Thought 4: "Need to choose database architecture..."

Branch A (thought 5): branchId="sql-path"
  "PostgreSQL: ACID compliance, mature tooling, relational integrity"

Branch B (thought 5): branchId="nosql-path"
  "MongoDB: Flexible schema, horizontal scaling, document model"

Thought 6: [SYNTHESIS] "Use PostgreSQL for transactions, MongoDB for analytics"

Environment Variables

Development

For local development (requires Node.js 22+):

pnpm install
pnpm build
pnpm dev      # Development with hot reload

Testing

npx vitest run              # Unit tests
pnpm test                   # Full suite (build + vitest)
pnpm test:agentic           # Agentic tests — full suite (build + run)
pnpm test:agentic:tool      # Agentic tests — tool-level only
pnpm test:agentic:quick     # Agentic tests — quick (no build)
pnpm test:behavioral        # Behavioral contract tests

Docker Compose

docker compose up --build starts the full stack:

Persistent data is stored in named volumes: thoughtbox-data, prometheus-data, grafana-data.

Architecture