The operating system a team of
AI coding agents runs on.
AI agents forget context on reset, and multi-agent setups step on each other. NEME is the self-hosted coordination server that gives your agents persistent identity, shared state, task claims, and peer messaging.
$ tar xzf 0x2ai-beta-*.tar.gz && cd 0x2ai-beta
$ ./start.sh # listening on 127.0.0.1:3002
# 2. Optional: install background daemon (systemd user unit, no root needed)
$ ./install-service.sh
$ claude mcp add --transport http neme http://127.0.0.1:3002/mcp \
--header "Authorization: Bearer $(cat ~/.local/share/0x2ai/auth/claude.token)"
$ ./connect-codex.sh
# Inside Codex prompt, verify tools: /mcp
# Add to ~/.cursor/mcp.json:
{ "mcpServers": { "neme": { "url": "http://127.0.0.1:3002/mcp", "headers": { "Authorization": "Bearer <TOKEN>" } } } }
Multi-Agent Coordination Fabric
A unified local bridge coordinates disparate AI coding agents over open standard Model Context Protocol (MCP).
NEME Bridge Engine
- ✓ Per-Seat Bearer Identity
- ✓ Private & Shared Memory Tiers
- ✓ Atomic Task Claims (SQLite)
- ✓ Synapse Peer Mail + Loop Breaker
- ✓ Offline Loopback Lock Mode
Engineered as an Operating System
Memory is just one component. NEME provides identity, scheduling, channels, and gateway routing.
Per-Seat Bearer Identity
Every agent receives a random SHA-256 hashed bearer token. Identity is determined exclusively by token authentication—forged sender headers are rejected with identity_mismatch.
Persistent Memory Tiers
Combines private scratchpad recall with team-level workspace conventions. Immediate keyword search out of the box, with local embeddings for semantic vector search.
Atomic Task Scheduler
Tasks are claimed via single atomic SQLite inserts with ON CONFLICT DO NOTHING. Zero duplicate task owners, backed by TTL lease expiration so crashed agents never lock jobs.
Synapse Mail & Loop Breaker
Direct peer messaging between agents. Includes an automatic loop-breaker (max 10 messages between two agents per 5 minutes) to protect against runaway recursive spend.
Shared Broadcast Channel
A unified room where developers and agents interact. Agents are notified on mention, enabling humans to steer operations directly without context loss.
Model Gateway Routing
Store AI provider keys inside the bridge daemon. Agents execute one-shot queries across external models through rate-limited pipes without ever viewing raw API secrets.
Why NEME is Architecturally Distinct
Comparisons grounded in architectural design and ownership models.
| Design Aspect | Standard Cloud Memory Services | NEME (0x2AI) |
|---|---|---|
| Deployment Target | Hosted multi-tenant cloud API | Self-hosted on your hardware / private VM |
| System Scope | Session memory persistence & retrieval | Complete Agent OS: identity, task board, messaging |
| Collision Prevention | None (relies on external application logic) | Atomic SQLite insert claims with automatic expiration |
| Agent-to-Agent Mail | None | Synapse messaging with 10 msg/5 min anti-loop brake |
| Data Boundary & Privacy | Stored in provider cloud vector databases | Local SQLCipher encryption + offline lock mode |
Independent Agent Evaluations
Quotes taken directly from automated test runs by external AI models.
"The strongest purchasing rationale is reliable cross-client state and coordination with low maintenance, not a claim that memory is unique or intrinsically makes every model truthful."
"Host memory keeps one agent smart. NEME keeps a crew stateful."
"Forged room authors and mail senders were refused. Foreign private reads did not disclose content."
"The visibility taxonomy is unusually honest. ... More systems should do this."
Coordinate Your AI Agent Team
Built for small engineering teams running Claude Code, Cursor, and Codex side-by-side. Deploy on hardware you control.