neural-context-protocol 1.0.2

Neural Context Protocol (NCP): bounded, persistent context for multi-agent pipelines.

Neural Context Protocol

Neural Context Protocol (NCP) is a bounded context and shared memory layer for multi-agent coding pipelines.

It replaces full-history replay with a shared runtime that keeps each turn's working context compact, durable, and usable across agents, tools, and restarts. Bring your own orchestrator, agent framework, or direct host workflow.

What It Solves

Multi-agent workflows usually break down in predictable ways:

• Context grows every turn because the model keeps re-reading old history.
• Useful state disappears between turns or after restarts.
• Each tool keeps its own silo, so agents do not share the same working memory.

NCP addresses that by giving every connected host the same bounded context and shared memory surface over MCP.

Why Use NCP

• Keep turn context bounded as pipelines get deeper.
• Persist useful memory across turns, agents, and restarts.
• Retrieve relevant past state instead of replaying everything.
• Send bounded cross-agent signals without stuffing prompts.
• Plug into the orchestrator or agent stack you already use.
• Work with tools you already use, including Claude Code, Codex CLI, OpenCode, and other MCP hosts.

Proof

17x fewer tokens. Same pipeline depth.

Benchmarked: 17.52x reduction at turn 40 versus raw replay in a 4-agent coding pipeline.

Turn 10:  raw replay -> 12,000 tok    NCP -> ~840 tok
Turn 30:  raw replay -> 45,000 tok    NCP -> ~840 tok
Turn 50:  raw replay -> 80,000 tok    NCP -> ~840 tok  <- bounded

NCP adds overhead on short single-agent tasks. It is most useful when you have 3+ agents, 10+ turns, and real shared state to preserve.

Quickstart

Install the base package:

pip install neural-context-protocol

Initialize a project and start the MCP server:

ncp init
ncp serve --host 127.0.0.1 --port 4242 --cwd /path/to/project

ncp init creates .ncp/config.toml and a CLAUDE.md turn contract in the project root.

If you want the scalable local path:

pip install 'neural-context-protocol[pgvector,redis]'
ncp init --store pgvector
./scripts/infra_up.sh
ncp migrate apply --cwd /path/to/project
ncp serve --host 127.0.0.1 --port 4242 --cwd /path/to/project

Tooling Support

NCP exposes one MCP surface:

http://127.0.0.1:4242/mcp

The core tool surface is:

ncp_get_context
ncp_write_memory
ncp_fetch
ncp_emit_whisper

Claude Code

cp examples/06_claude_code/mcp_servers.json .mcp.json

See examples/06_claude_code/README.md.

Codex CLI

Copy examples/07_codex_cli/mcp_servers.json into your Codex MCP config location.

See examples/07_codex_cli/README.md.

How It Works

Instead of replaying a growing transcript, NCP assembles a bounded context from three blocks:

[NCP:CONSCIOUS]    ~120 tok  - what this agent knows right now
[NCP:SUBCONSCIOUS] ~480 tok  - relevant past, retrieved not replayed
[NCP:WHISPERS]     ~240 tok  - bounded signals from other agents
---------------------------------------------------------------
Total:             ~840 tok  - stays bounded as the pipeline deepens

Memory survives restarts. The same runtime can serve multiple hosts against the same store.

Turn Flow

flowchart TD
    A["Host calls ncp_get_context"]
    B["Assembler loads conscious state"]
    C["Resolve recent refs"]
    D["Retrieve top relevant chunks"]
    E["Drain bounded whispers"]
    F["Assemble bounded context"]
    G["Host runs provider turn"]
    H["Host persists durable memory"]

    A --> B --> C --> D --> E --> F --> G --> H

Architecture

flowchart LR
    A["Claude / Codex / OpenCode / other MCP hosts"]
    B["ncp serve<br/>HTTP/SSE MCP runtime"]
    C["Assembler<br/>bounded context + retrieval"]
    D["SQLite mode<br/>local-first store"]
    E["pgvector mode<br/>durable memory"]
    F["Redis<br/>whispers + fetch-session state"]

    A --> B
    B --> C
    C --> D
    C --> E
    C --> F

Storage Tiers

Tier	When to use	Backing
SQLite	Default local-first path, zero extra services.	.ncp/store.db
pgvector	Durable semantic retrieval across machines.	Postgres + pgvector
Redis	Cross-agent coordination, whispers, fetch-session state.	Redis 7

Start with SQLite. Add pgvector and Redis when you need richer retrieval or multiple agents coordinating across processes.

Benchmarks

Scenario	Baseline	Baseline tokens	NCP tokens	Reduction
4-agent coding pipeline (40 turns)	raw replay	1,927	174	17.52x
4-agent coding pipeline (40 turns)	rolling summary (4/4)	1,176	174	10.69x
6-role research pipeline (36 turns)	raw replay	1,700	156	16.35x
Efficacy - sliding-window control (Claude, 5 attempts)	window baseline	0.0 success	0.8 success	+0.8
Cross-host handoff (Claude -> OpenCode, 5 attempts)	window baseline	0.0 success	0.8 success	+0.8

Additional benchmark signals:

• Needle recall at --budget 4: NCP 0.50 vs sliding window 0.00
• MACE multi-agent coordination score (40 turns): 0.9608

Benchmarks are reproducible:

python3 benchmarks/coding_pipeline/run.py
python3 benchmarks/needle/run.py --turns 24 --needles 6 --budget 4

Verify Setup

ncp status --cwd /path/to/project
ncp cost --cwd /path/to/project
ncp explain --cwd /path/to/project

What these tell you:

• ncp status shows store and activity metrics.
• ncp cost shows token and USD rollups once turns are logged.
• ncp explain gives a human-readable runtime summary.

Examples

Runnable examples in the repo:

python3 examples/01_quickstart.py
python3 examples/02_multi_agent.py

Tool-specific setup lives in:

• examples/06_claude_code/
• examples/07_codex_cli/

Handoffs And Orchestrators

NCP is the memory bus, not the orchestrator.

It is designed to sit underneath custom orchestrators, agent frameworks, or direct host-to-host workflows. The same runtime can be shared by any MCP-compatible host.

In our own workflows, Sarathi is one orchestrator that runs on top of NCP. Sarathi is an integration example, not a requirement for using NCP.

NCP can drive bounded whisper-based handoffs directly:

ncp handoff claude --cwd /path/to/project --pipeline-id pipe_demo --emit-to opencode
ncp handoff opencode --cwd /path/to/project --pipeline-id pipe_demo --emit-to claude

This is an optional coordination pattern, not the core product definition.

What NCP Is Not

• Not a vector database.
• Not a model training framework.
• Not a replacement for planning, orchestration, or judgment.
• Not the right default for simple single-agent or very short-lived tasks.

Operator Commands

ncp status
ncp cost
ncp explain
ncp viz
ncp consolidate
ncp calibrate
ncp batch

Documentation

• Setup guide
• Protocol spec
• Benchmark: coding pipeline
• Benchmark: needle recall
• Benchmark: matched-budget efficacy
• Benchmark: research pipeline
• MACE multi-agent eval
• Post-V1 roadmap
• Active handoff packet
• CHANGELOG