Integrate

The gateway is the switchboard; this page is about what plugs into it.

Pair with a neuralOS instance (verified data answers)

A neuralOS instance is a verified data-answer runtime: the model picks from a closed menu of probes, arguments are caged, weak matches are refused, and every query is audited. Served behind an agentd, it becomes a mesh capability — peers delegate data questions and get verified answers or an honest refusal.

# 1. serve the instance with token auth
neuralosd serve --instance-dir ./finacle-users --port 8877 \
                --auth-token "$NEURALOSD_TOKEN"

# 2. expose it as the agent's toolbox
neuralosd openapi ./finacle-users      # OpenAPI + MCP manifest

# 3. attach the agentd whose agent uses that toolbox
synapse-agentd -gateway ws://127.0.0.1:3000/ws/v2/agent \
  -agent-id finacle-users-agent \
  -provider-key "$LLM_KEY" -workdir /srv/agentd-work
# agent config: MCP server → http://127.0.0.1:8877

What crosses the mesh afterwards: verified answers or typed refusals — never database credentials, never invented numbers. Two correlated audit trails record the whole chain (gateway ledger + instance audit).

Why this pairing matters:

Problem	How the pair answers it
"Can another team query our data with an LLM?"	Only a menu of caged questions is exposed; no DB access
Hallucinated figures in agent workflows	Worst case is a typed refusal — an invented number is structurally excluded
Compliance forensics	Gateway ledger + instance audit reconstruct question → probe → data
Data residency	The instance never leaves your premises; only answers cross the mesh

Webhooks — know when mesh tasks finish

Async tasks a peer creates on your edge can notify your systems:

synapse-gateway -mesh-enabled … \
  -mesh-task-webhook https://ops.example.com/hooks/task-done \
  -mesh-task-executor-webhook https://ops.example.com/hooks/task-input

Notifications are signed and POSTed on completion (or when a task you execute pauses for input — the task and its question are the payload). Verify signatures with the edge identity before acting on them.

MCP and AI products

The gateway's directory + gates are agent-agnostic. Any MCP-capable client or AI product can consume an edge's capabilities; the neuralOS pairing above is the pattern for making those capabilities hallucination-proof.

Multi-organisation meshes

1. Each org runs its own gateway (data stays behind each edge)
2. Share one NATS server (or cluster) as the discovery mesh
3. Each edge: -mesh-enabled + credentials + -mesh-capabilities
4. Verification posture for regulated meshes: -mesh-verify-mode require -mesh-trust-on-first-use=false -mesh-trusted-peers sha256:…,sha256:…
5. Bound escalation: -mesh-require-verified-invoke (default on), -mesh-invoke-operations, -mesh-invoke-timeout, -mesh-skill-proxy-targets

Discovery is public by design; invocation is not.