ASTRA — Unified Research Lab + MCP Server

Autonomous Sentient Thoughtful Reasoning Agent

Production-grade Model Context Protocol server exposing the ASTRA bio-hybrid neuromorphic simulation pipeline to AI assistants. Built with the official @modelcontextprotocol/sdk, it integrates a layered SNN LIF+STDP engine, consciousness proxy assessment, bio-computing platform telemetry, and an IRB ethics monitor — all queryable as MCP tools, resources, and prompts from Claude Desktop, Cursor, VS Code, and any MCP-compatible client.

🆕 v2.4 — Second-Order (Self-Evidencing) Loop

ASTRA v2.4 adds the second-order loop: the update organ the double-loop architecture (Legros 2026, §3.2) requires. v2.2 produced the signal of self-evidencing (world-model surprise) but not the layer that observes and corrects the system's own predictive capacity. v2.4 supplies it as a native TypeScript port of the second-order fragments of the_consciousness_ai:

• Meta-learning velocity (MetaLearningModule) — convergence vs novelty from RPE-variance dynamics

• Curiosity / intrinsic reward (RNDCuriosity) — the EFE epistemic-value (exploration) term

• Capability model (DirectExperienceLearner) — action → expected-valence agency map

• Meta-consciousness and developmental tracking — second-order self-monitoring over time

6 new MCP tools (tcai_second_order, tcai_meta_learning, tcai_capability_model, tcai_curiosity, tcai_metaconsciousness, tcai_development). A Friston-exact pymdp counterpart is provided as an optional Python bridge. See SECOND-ORDER-LOOP-INTEGRATION.md.

Related MCP server: ASTRA — Unified Research Lab + MCP Server

🆕 v2.2 — the_consciousness_ai (ACM) Integration

ASTRA v2.2 integrates tlcdv/the_consciousness_ai — the Artificial Consciousness Module research codebase — at two levels:

• Native TypeScript port (src/engine/tcai/): Global Neuronal Workspace with sigmoid ignition & reverberation, Kuramoto/AKOrN oscillatory binding, PAD emotional processing & reward shaping, attention-gated emotional memory, self-representation core + attention schema, and a metrics suite (GNW · Effective Information · Φ̃-RIIU) — all fed live from the SNN/world-model state and exposed as 8 new MCP tools (tcai_cycle, tcai_workspace_state, tcai_emotion_appraise, tcai_memory_store, tcai_memory_retrieve, tcai_self_model, tcai_metrics, tcai_reset).

• Full vendored Python codebase (python/the_consciousness_ai/, 215 files): the complete upstream ACM project for reference and PyTorch-based reproduction.

See TCAI-INTEGRATION.md for the complete Python → TypeScript mapping and architecture coupling. All consciousness-related metrics remain computational proxies, not measurements.

🆕 v2.2 — FinalSpark NeuroPlatform v2 Integration

ASTRA v2.2 also integrates the FinalSpark NeuroPlatform v2 wetware control API — the closed-loop interface to living neural organoids on a 128-electrode MEA — at two levels:

• Native TypeScript port + biophysical simulator (src/engine/neuroplatform.ts): faithful port of the NeuroPlatform controller surface (StimParam with charge-balance checking, IntanController, TriggerController, DatabaseController, CameraController) backed by a seeded OrganoidMEA model — exposed as 9 new MCP tools (np_status, np_configure_stim, np_send_trigger, np_count_spikes, np_query_spike_count, np_query_spike_events, np_query_triggers, np_camera_capture, np_closed_loop). The MEA's 128 electrodes couple one-to-one with the ASTRA SNN's 128 neurons.

• Live Python bridge (python/neuroplatform/astra_np_bridge.py): runs a homeostatic closed loop against the physical platform via the genuine neuroplatformv2 SDK, streaming couplings to ASTRA over JSON-RPC.

• Standalone dashboard (dashboard/ASTRA-NeuroPlatform-Dashboard.html): live MEA raster, spike scope, StimParam editor with charge-balance readout, trigger generator and closed-loop telemetry.

See NEUROPLATFORM-INTEGRATION.md for the complete API → TypeScript mapping. With no hardware attached the server runs in simulate mode (deterministic biophysical model), not living-tissue measurements.

FinalSpark (800K neurons) ──┐
Cortical Labs CL1 ──────────┼─→ Spike Encoders → SNN (LIF+STDP, 128 neurons) → ACM Proxies
Koniku Kore ────────────────┘         │                    │
                                      │              ┌─────┴─────┐
                                      │              │  Φ̃  GW̃  PAD̃  │
                                      │              └─────┬─────┘
                                      ├─→ TCAI/ACM Layer (GNW · AKOrN · PAD · Φ̃-RIIU · EI)
                                      ├─→ NeuroPlatform v2 Bridge (MEA ↔ SNN · StimParam · closed loop)
                                      ├─→ Ethics IRB Monitor (mode-aware)
                                      └─→ MCP Server (48 tools · 11 resources · 8 prompts)

Note on data mode: In the default sim mode, all bio-platform data is synthetically generated. The server is designed to connect to live platforms in live mode, but this requires hardware access and appropriate IRB approval.

What's New in v2

• Layered SNN architecture: Configurable feed-forward + recurrent topology (default: 32→64→16→16 = 128 neurons) replacing the flat random network

• Event-driven STDP: O(spikes × fan-out) instead of O(N²) per timestep

• Ring buffer: O(1) spike history eviction replacing O(n) Array.shift()

• Sparse weight storage: Adjacency lists instead of dense N×N matrix

• Honest ACM naming: Proxies clearly labelled as integrationProxy, broadcastProxy, arousalProxy with methodological basis strings — no false IIT/GWT/PAD claims

• Bounds-checked parameters: set_parameter rejects implausible values (NaN, Infinity, out-of-range)

• Mode-aware ethics: Reports distinguish simulated vs live data with explicit disclaimers

• CI pipeline: GitHub Actions for build, test, and Docker smoke-test

• Repo hygiene: dist/ excluded from VCS, .gitignore added, deployment script removed

Quick Start

git clone https://github.com/christophejlegros-lgtm/ASTRA-Unified-ResearchLab-MCP-v2.4.git
cd ASTRA-Unified-ResearchLab-MCP-v2.4

# Install & build
npm install
npm run build

# Run (stdio — for Claude Desktop / Cursor)
node dist/index.js

# Or dev mode (no build needed)
npm run dev

Transports

Client Configuration

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "astra": {
      "command": "node",
      "args": ["/absolute/path/to/dist/index.js"],
      "env": { "ASTRA_LOG_LEVEL": "info" }
    }
  }
}

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "astra": {
      "command": "node",
      "args": ["/absolute/path/to/dist/index.js"]
    }
  }
}

VS Code

Add to .vscode/settings.json:

{
  "mcp": {
    "servers": {
      "astra": {
        "type": "stdio",
        "command": "node",
        "args": ["${workspaceFolder}/dist/index.js"]
      }
    }
  }
}

Docker (remote SSE + HTTP)

docker compose up -d
# SSE: http://host:9002/sse
# HTTP: http://host:9003/mcp

MCP Tools (41)

All tools declare MCP annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) and human-readable titles. Core tools below; see TCAI-INTEGRATION.md for the 8 tcai_* tools and NEUROPLATFORM-INTEGRATION.md for the 9 np_* tools.

MCP Resources (10)

MCP Prompts (7)

Pre-built workflow templates that orchestrate multi-tool sequences:

Architecture

.github/workflows/
└── ci.yml                # GitHub Actions: build, test, Docker smoke-test

src/
├── index.ts              # stdio transport entry point
├── sse-server.ts         # SSE transport (Express)
├── http-server.ts        # Streamable HTTP transport (Express)
├── server.ts             # MCP server factory (48 tools + 8 prompts + 11 resources)
│   ├── server-wm-tools.ts    # World Model JEPA tools (6 tools + 2 resources + 1 prompt)
│   ├── server-sensor-tools.ts # Multimodal sensor tools (6 tools + 1 resource + 1 prompt)
│   ├── server-tcai-tools.ts  # TCAI/ACM tools (15 tools + 2 resources + 2 prompts incl. second-order loop + halting criterion)
│   ├── server-neuroplatform-tools.ts # NeuroPlatform v2 tools (9 tools + 1 resource + 1 prompt)
├── engine/
│   ├── state.ts          # Reactive state store + parameter bounds registry
│   ├── snn.ts            # Layered SNN LIF+STDP engine (Map-indexed sparse weights, event-driven)
│   ├── acm.ts            # Consciousness proxy module (Φ̃ + GW̃ + PAD̃)
│   ├── ethics.ts         # IRB ethics monitor (mode-aware, biomarker thresholds)
│   ├── world-model.ts     # JEPA World Model engine (LeWM adapted)
│   ├── wm-simulation.ts   # WM simulation manager (replay buffer, auto-train)
│   ├── multimodal-sensors.ts # V-JEPA 2 + A-JEPA + Koniku + fusion
│   ├── neuroplatform.ts      # FinalSpark NeuroPlatform v2 port + OrganoidMEA simulator
│   └── simulation.ts     # Background tick loop
└── utils/
    └── logger.ts         # Structured logging (pino → stderr)

tests/
├── astra.test.ts         # Unit tests: state, bounds, SNN, ACM, ethics, security
├── world-model.test.ts   # World Model: encoder, predictor, SIGReg, CEM, surprise
├── wm-simulation.test.ts # WM simulation: buffer, training, planning, lifecycle
├── multimodal-sensors.test.ts # Sensors: V-JEPA, A-JEPA, Koniku, fusion, pipeline
├── tcai.test.ts          # TCAI/ACM: binding, GNW, memory, emotion, self-model, metrics
├── neuroplatform.test.ts # NeuroPlatform: StimParam, OrganoidMEA, controllers, bridge
└── integration.test.ts   # Client SDK integration: tools, resources, prompts, workflow

configs/                  # Ready-to-use client configurations

Extracted to separate repositories: The v1 HTML dashboard (4 669 lines) and the legacy Node.js bridge config have been removed from this repo to keep it focused on the MCP server. See ASTRA-Unified-ResearchLab-MCP- for the original dashboard.

SNN Engine

Layered LIF+STDP — Configurable layered architecture. Default: 32 (input) → 64 (hidden_1) → 16 (hidden_2) → 16 (output) = 128 neurons.

Connectivity: feed-forward between adjacent layers (30%) + sparse recurrent within layers (10%). Weights stored as sparse adjacency lists, not dense matrices.

Biophysical parameters: τ_m = 20ms, V_th = −50mV, V_reset = −70mV, refractory = 2ms. Background noise range [10, 22] mV produces ~2 spikes/step at steady state with all neurons active. STDP: A+ = 0.01, A− = 0.012, τ± = 20ms, event-driven (processes only spiking neurons per timestep).

The SNN topology resource (astra://snn/topology) dynamically reports the actual engine configuration, including layer sizes, synapse count, connectivity parameters, and weight storage type (Map-indexed sparse adjacency lists).

ACM — Consciousness Proxy Module

⚠ Methodological disclaimer: The metrics below are computational proxies inspired by the referenced theories. They are not faithful implementations. See source code comments for full details.

Composite score: ACM = α·Φ̃ + β·GW̃ + γ·PAD̃ (default: α=0.40, β=0.35, γ=0.25)

Ethics IRB Monitor

IRB compliance level N3 (100K–1M neurons). Four biomarkers with three-state classification.

Mode-aware: In sim mode, reports include explicit disclaimers that data is synthetic and irbRequired is false. In live mode, DISTRESS triggers mandatory IRB notification.

Parameter Bounds

The set_parameter tool validates all numeric inputs against a bounds registry to prevent injection of absurd values (negative percentages, Infinity, NaN). Bounds are defined per parameter path — see src/engine/state.ts for the complete registry.

Testing

# Full suite
npm test

# Unit tests only
node --import tsx --test tests/astra.test.ts

# Integration tests only (Client SDK)
node --import tsx --test tests/integration.test.ts

# TCAI / NeuroPlatform suites only
npm run test:tcai
npm run test:np

# MCP Inspector
npm run inspect

Full suite: 204/204 passing (188 prior + 16 second-order loop), 0 TypeScript errors (strict, Node16 ESM).

Development

npm run dev        # stdio (no build)
npm run dev:sse    # SSE on :9002
npm run dev:http   # HTTP on :9003
npm run watch      # TypeScript watch mode

Environment Variables

Scaling Notes

The default 128-neuron configuration is designed for interactive demonstration. To scale toward the aspirational 256→512→256→128 (1 152 neurons) architecture:

1. Pass custom layers to SNNEngine: new SNNEngine({ layers: [{ name: 'input', size: 256 }, ...] })

2. Event-driven STDP scales as O(spikes × average fan-out), not O(N²)

3. Map-indexed adjacency lists provide O(1) weight lookup per synapse

4. Sparse storage keeps memory proportional to actual synapses (~18 KB at 128 neurons vs 64 KB dense)

5. Consider increasing intervalMs in the simulation loop for larger networks

6. For >10K neurons, a Rust/WASM or Lava SDK backend is recommended

License

MIT — © 2026 Christophe Jean Legros, Geneva

Assistance Multi IA · Assistant-Multi-AI@proton.me

References

• Model Context Protocol · Spec 2025-11-25

• MCP TypeScript SDK

• FinalSpark · Cortical Labs · Koniku

• Intel Lava / Loihi 2

• Gerstner & Kistler (2002) "Spiking Neuron Models"

• Tononi (2004) "An information integration theory of consciousness" — BMC Neuroscience

• Baars (1988) "A Cognitive Theory of Consciousness" — Cambridge University Press

• Mehrabian (1996) "Pleasure-Arousal-Dominance: A General Framework" — Current Psychology