Platform Architecture & Tech Stack
Version 1.9·August 2026·Status: OS-level canonical 12 platform components + 1 conditional·Multi-domain OS substrate
Overview
Synedrix OS is a multi-agent operating system built primarily on open-source infrastructure platforms. Each platform has a distinct, non-overlapping responsibility. Nearly all components are self-hostable and can be operated within a customer’s own infrastructure — with one documented exception, below.
The stack is a reusable operating-system substrate for business execution. Domain packs run on top of it without modifying its core abstractions.
Every component in this stack is open source and self-hostable, except Hookdeck (the inbound Event Gateway). Hookdeck is a proprietary, closed-source managed cloud service with no documented self-hosted, on-prem, or private-cloud deployment — see Hookdeck platform reference: Deployment and licensing. Hookdeck Outpost, its egress counterpart, is a separate open-source (Apache 2.0) codebase and is genuinely self-hostable; the two are frequently conflated because they share a vendor name. This exception is material to Principle 11 (Open by default, sovereign by design), which every other component in this stack satisfies, and to Mode 3 (air-gapped). The exception is accepted (ADR-OS-013): Hookdeck is Mode 3’s single declared external dependency — see Hookdeck platform reference: Mode 3 is the hard problem for the decision.
Deployment modes
Three canonical deployment modes govern how the stack is operated. The deployment mode is declared at the business instance level and constrains model routing for all agents in that instance.
Mode 1 — Cloud-connected
All platform components run self-hosted on the customer’s infrastructure. Model inference uses external cloud APIs.
When to use: organisations comfortable with standard cloud data processing agreements; development and early production environments.
Mode 2 — Self-hosted hybrid
All platform components run self-hosted. Model inference is split at the agent level — some agents use external cloud APIs, others use local models.
When to use: organisations that want local inference for sensitive workflows and cloud models for lower-sensitivity tasks.
Mode 3 — Air-gapped (fully local)
All platform components run self-hosted except the inbound Hookdeck Event Gateway, Mode 3’s single declared external dependency. No outbound connections or external API calls of any kind. No business data leaves the customer’s infrastructure; inbound event payloads transit Hookdeck before delivery. See the Deployment Model Spec.
- No cloud TTS provider URI in any agent runtime policy.
When to use: regulated industries, government, defence, any environment where data sovereignty is non-negotiable.
Platform Responsibility Matrix
What each of the 12 platforms in the Synedrix OS tech stack owns, and where it sits in the layer model.
| Platform | Primary role | Access pattern | Latency profile | OS layer | Self-hostable | Required in |
|---|---|---|---|---|---|---|
| n8n | Orchestration & workflow control | Read/Write | Low-latency execution | Orchestration | Yes (Apache 2.0) | All modes |
| LangGraph | Agent runtime: stateful/cyclic/multi-agent reasoning | Invocation R/W | Reasoning-bound | Agent runtime | Yes (MIT) | All modes |
| Hookdeck | Inbound event ingress & buffering | External write | Sub-second | Ingress gateway | No (proprietary, managed cloud only) | All modes — accepted exception in Mode 3 (decision) |
| Hookdeck Outpost | Outbound notification delivery | Write from workflows | Sub-second | Egress gateway | Yes (Apache 2.0) | All modes |
| Redis | Hot state, pub/sub, deduplication | Fast R/W | Sub-10 ms | Coordination bus | Yes (BSD) | All modes |
| Supabase PG17 | Canonical relational truth & config | Standard R/W | Low tens of ms | Transactional state | Yes (Apache 2.0) | All modes |
| QuestDB | Time-series data engine | High-throughput R/W | Low-latency | Temporal data layer | Yes (Apache 2.0) | All modes |
| Neo4j | Graph lineage & relationship context | Async write/read | Low to medium | Context graph | Yes (Community — GPL3) | All modes |
| Qdrant | Similarity search & vector retrieval | Async write/query | Low-latency | Semantic memory | Yes (Apache 2.0) | All modes |
| Langfuse | LLM traces, prompts, evals, token/cost | Async write/read | Async to low-latency | LLM engineering | Yes (MIT) | All modes |
| PostHog | Behavioral analytics, feature flags, experiments | Async write/read | Async | Product/decision analytics | Yes (MIT) | All modes |
| Grafana + Prometheus | Infrastructure observability & alerting | Scrape + read | N/A | Operational monitoring | Yes (AGPL/Apache) | All modes |
| Ollama / vLLM | Local model inference | Inference R/W | Hardware-dependent | Model inference | Yes (MIT / Apache 2.0) | Modes 2 & 3 only |
1. n8n — Orchestration Core
Platform reference: n8n — what it is, what it is normally used for, and how and why Synedrix OS uses it.
The orchestration kernel of Synedrix OS.
What it owns: workflow execution, agent invocation routing, trigger and action composition, retry and timeout management, approval routing, sub-workflow calls, and the execution graph for all domain pack workflows.
Layer 3 against Layer 4, as of v1.5: As of ADR-OS-001, n8n is the Orchestration Layer (Layer 3) only. Complex agent reasoning — stateful, cyclic, or multi-agent patterns (
iterative,consensus,reflect-and-revise,hierarchical) — runs in the LangGraph Agent Runtime (Layer 4, Platform 15), invoked by n8n as a governed step. Simple acyclic agents may still run as n8n AI Agent nodes. See the Orchestration & Runtime Strategy for the routing rule and boundary contract.
Self-hosting: n8n is open source under Apache 2.0. Full feature set available in self-hosted. Runs as Docker container or Kubernetes deployment.
2. Hookdeck — Inbound Event Gateway
Platform reference: Hookdeck — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: inbound event validation, schema enforcement, buffering, replay, and fan-out.
Self-hosting: none. The Hookdeck Event Gateway is a proprietary, closed-source managed cloud service — no self-hosted, on-prem, or private-cloud deployment is documented. This is the one component in the stack that does not satisfy Principle 11 as written; the exception is accepted and recorded in ADR-OS-013, and it makes Hookdeck the single declared external dependency of Mode 3 (air-gapped). See Hookdeck platform reference: Deployment and licensing for the full analysis.
3. Hookdeck Outpost — Outbound Notification Gateway
Platform reference: Hookdeck Outpost — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: durable outbound event and webhook delivery from Synedrix OS workflows to external systems. Ensures delivery guarantees for outbound notifications regardless of downstream availability.
Self-hosting: Outpost is open source under Apache 2.0 and self-hosts independently of Hookdeck (a separate codebase and licence) via Docker, Kubernetes, or Railway. Its dependencies — Redis, PostgreSQL, and a message queue — are already in this stack. It is not “self-hosted alongside Hookdeck”: Hookdeck itself is not self-hostable at all. See Hookdeck Outpost platform reference: Deployment and licensing.
4. Redis — Coordination Bus & Hot State
Platform reference: Redis — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: hot state, pub/sub broadcast, idempotency keys, rate-limit counters, heartbeats, and deduplication windows.
Self-hosting: Redis open source under BSD 3-Clause. Redis 7+ recommended.
5. Supabase PG17 — Canonical Relational Truth
Platform reference: Supabase PG17 — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: business entities, workflow state snapshots, policies, approvals, audit trails, operator configuration, domain-pack data models, and agent definitions.
Self-hosting: Supabase is open source under Apache 2.0. Self-hosted version runs the full Postgres 17 stack.
6. QuestDB — Time-Series Data Fabric
Platform reference: QuestDB — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: high-volume time-series records, materialised temporal aggregations, trend and seasonality query support, historical event windows.
Self-hosting: QuestDB is open source under Apache 2.0.
7. Neo4j — Context Graph & Lineage Layer
Platform reference: Neo4j — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: workflow and outcome lineage, agent influence and correlation networks, entity relationship graphs, cross-system dependency context.
Self-hosting: Neo4j Community Edition is open source under GPL3. Enterprise features require Neo4j Enterprise licence.
8. Qdrant — Semantic Memory & Similarity Retrieval
Platform reference: Qdrant — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: semantic retrieval collections, similarity search, metadata-aware filtered vector queries, pattern recall.
Self-hosting: Qdrant is open source under Apache 2.0. Runs as single binary or Docker container.
9. Langfuse — LLM Engineering Layer
Platform reference: Langfuse — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: prompt versioning, LLM/agent trace capture, token and cost attribution, model-call latency and error visibility, evaluations and quality scoring, nested trace debugging.
Self-hosting: Langfuse is open source under MIT. Self-hosted version stores all data locally with no external telemetry.
Mode 3 note: In air-gapped deployments, Langfuse runs fully locally. Traces are captured from local STT/TTS and LLM inference endpoints using standard API formats.
10. PostHog — Behavioral Analytics & Feature Flags
Platform reference: PostHog — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: workflow analytics, feature flags, A/B experiments, business decision analytics.
Self-hosting: PostHog is open source under MIT. Self-hosted version stores all analytics data locally.
11. Grafana + Prometheus — Operational Monitoring
Platform reference: Grafana LGTM Stack — what it is, what it is normally used for, and how and why Synedrix OS uses it.
What it owns: infrastructure health, service dashboards, alerting, and SLOs.
Self-hosting: Grafana (AGPL 3.0), Prometheus (Apache 2.0). Both run via Docker or Kubernetes.
12. Ollama / vLLM — Local Model Inference (Modes 2 & 3 only)
What it owns: local model inference for LLM-backed agents. Exposes an OpenAI-compatible API so Langfuse tracing, evaluation, and cost tracking integrations work without modification.
Self-hosting: Ollama (MIT), vLLM (Apache 2.0).
Mode 3 note: In air-gapped deployments Ollama/vLLM is the only inference layer. GPU hardware is recommended for production workloads; CPU inference is functional but may not meet latency requirements for interactive use.
13. LangGraph — Agent Runtime Core
Platform reference: LangGraph — what it is, what it is normally used for, and how and why Synedrix OS uses it.
The dedicated Agent Runtime (Layer 4) for stateful, cyclic, and multi-agent reasoning. Pairs with n8n at the orchestration core: n8n orchestrates (Layer 3), LangGraph reasons (Layer 4). Introduced by ADR-OS-001.
What it owns:
- Stateful agent graphs with checkpointing and durable execution.
- The cyclic/multi-agent execution patterns:
iterative,consensus,reflect-and-revise,hierarchical. - Human-in-the-loop interrupts that suspend and resume a graph at a checkpoint (wired to the OS HITL Gateway).
- Tool-calling reasoning loops, built with the LangChain library (model abstraction, tool calling, structured output, RAG over Qdrant, MCP adapters). LangChain is a library inside this platform, not a separate platform row.
What it does not own: workflow triggers and business sequencing (n8n, Layer 3), policy decisions (Policy Guard os-agt-002), tool authorization resolution (orchestration computes the active tool set and passes it in), LLM tracing/eval (Langfuse, Layer 10), canonical state (Supabase).
Boundary contract: LangGraph is invoked by n8n as a governed step over HTTP / MCP / A2A. The invocation carries correlation_id, causation_id, task input, the resolved active_tool_set, and deployment_mode; the runtime returns the canonical Outcome Envelope and emits os.outcome.* events. The executor rejects any tool call outside the resolved active set. Full governance bridge: Orchestration & Runtime Strategy §5.
Observability: LangGraph and LangChain emit traces to Langfuse (OpenTelemetry / native exporter) — the same Layer 10 platform used by n8n agents. LangSmith is not part of the canonical stack (proprietary SaaS; violates Principle 11 and Mode 3; duplicates Langfuse). It MAY be used only as an optional developer-time convenience in cloud-connected Mode 1.
Deployment modes: All modes. Self-hostable and runs against local inference (Ollama/vLLM) in Modes 2 and 3, so the air-gapped invariants hold.
Self-hosting: LangGraph is open source under MIT. Runs as a Python service (Docker/Kubernetes); LangGraph’s durable-execution server is self-hostable.
Self-hosted stack integration: The Synedrix OS self-hosted stack ships a reference LangGraph agent service (langgraph-agent) defined in langgraph-compose.yml in the synedrix-os-stack-self-hosted repository. It exposes the governed HTTP invocation endpoint at port 8765 on the synedrix_core Docker network. The service connects to SearXNG (read-only search tool), Langfuse (trace emission), and the LiteLLM proxy (model routing). The POC workflow contract governing this service is specified in POC: n8n → LangGraph Governance Bridge Workflow (WF-OS-POC-001).
Licence: MIT (LangGraph), MIT (LangChain)
OS layer: Agent runtime layer
Layered Truth Summary
Synedrix OS separates different kinds of truth across the stack:
This is the substrate axis — which store holds which kind of truth. It is orthogonal to the tier axis (what kind of memory a record is), defined in Memory Fabric Spec §3. Neither may be inferred from the other.
| Truth class | Platform | Notes | Memory tiers it backs |
|---|---|---|---|
| Canonical relational | Supabase PG17 | Business entities, workflow state, approvals, audit trails. Write-authoritative for every audit-bearing tier | episodic, semantic, procedural, profile, relational, graph (lineage) |
| Graph / lineage | Neo4j | Workflow and outcome lineage, agent influence networks, cross-session entity graphs | graph (entities, via the NAMS client) |
| Temporal | QuestDB | High-volume time-series records, materialized aggregations, event windows | temporal |
| Semantic / similarity | Qdrant | Vector representations for similarity search and case recall. The single vector substrate — no second vector store enters Layer 7 | index for episodic, semantic, procedural, profile |
| Ephemeral / hot state | Redis | Operational state, pub/sub, dedup windows, idempotency keys. Not a source of truth | working |
| LLM eval / observability | Langfuse | Agent traces, prompt versions, token and cost attribution | none — traces are not a system of record |
Libraries adopted, and components deliberately not adopted
The stack’s defining constraint is that each platform holds one non-overlapping responsibility. That constraint decides adoption questions more often than capability does — the same axis on which LangSmith was rejected in ADR-OS-001, applied to storage rather than licensing.
| Component | Position | Reason |
|---|---|---|
| LangChain | Adopted as a library inside the Layer 4 runtime | Not a platform row. Model abstraction, tool calling, RAG, MCP adapters. |
NAMS — neo4j-labs/agent-memory, Apache 2.0 | Adopted as a library over the existing Neo4j | Runs on infrastructure the stack already operates: a schema plus an async client, in self-hosted bolt mode. Supplies the graph tier’s POLE+O ontology and the reasoning-to-entity lineage edges. Its own vector index is scoped to entity resolution so Qdrant remains the single vector substrate. See ADR-OS-002 §3.2. |
NAMS hosted service (memory.neo4jlabs.com) | Not adopted | Proprietary SaaS. Breaks Principle 11 and Mode 3 — the LangSmith precedent. Self-hosted bolt only, in every deployment mode. |
| EverOS model — episode → consolidation → profile/skill distillation | Adopted as the canonical memory lifecycle | Implemented on the existing fabric; see Memory Fabric Spec §5. |
| EverOS runtime — Markdown + SQLite + LanceDB, Apache 2.0 | Not adopted | LanceDB duplicates Qdrant, SQLite duplicates Supabase, and Markdown-as-truth contests Supabase’s write-authoritative role. Adopting it would add a fourteenth and fifteenth datastore to a stack whose central claim is non-overlapping responsibilities. Evaluated under the governed lab lane as LAB-001 in synedrix-os-stack-self-hosted, not as an unmanaged reference. |
| LangSmith | Not adopted | Proprietary SaaS; duplicates Langfuse. See ADR-OS-001 §3.3. |
Open item. NAMS is an experimental Neo4j Labs project with community support and no SLA, so the dependency must be version-pinned and carry an exit path. Its Neo4j 5.20+ vector-index requirement is satisfied by Community Edition (verified against the Operations Manual per-edition feature table), so the
graphtier is Mode 3 ready on Community. See Neo4j platform reference.
Interoperability Position
- Synedrix OS is standards-aware at the interoperability layer.
- MCP is the preferred protocol family for agent-to-tool and agent-to-context integration.
- A2A is the preferred protocol family for cross-system agent interoperability.
- In Mode 3, all MCP-connected tools and A2A-connected agents must be locally accessible. No external network dependency is permitted.
Why the Stack Holds at OS Level
This stack works for Synedrix OS because it is designed around reusable, open-source infrastructure concerns that span all business domains. Every component is self-hostable, open source, and independently operable. A domain pack needing infrastructure of its own — the Trading pack’s broker-execution adapters, for instance — supplies and owns it at Layer 11 rather than growing the OS. By expressing the architecture as a set of well-defined platform roles, Synedrix OS can:
- Support different domain packs on a single governed substrate
- Support all three deployment modes without architectural change
- Remain open to standards (MCP, A2A) while preserving core governance
Versioning
- Document: Synedrix OS Tech Stack
- Version: 1.9
- Applies to: Synedrix OS core platform and all domain packs
- Supersedes: v1.8
- Status: Canonical OS-level reference
Related Documents
Canonical Reference
- Synedrix OS - Consolidated Canonical Reference — Consolidated synthesis; §8 summarizes the platform stack defined here.
Orchestration and Runtime
- Synedrix OS ADR-OS-001 — Agent Runtime Split — The decision that added LangGraph (Platform 15) and repositioned n8n.
- Synedrix OS Orchestration & Runtime Strategy — n8n↔LangGraph boundary, routing rule, and governance bridge.
Deployment and Principles
- Synedrix OS - Deployment Model Spec — Three canonical deployment modes; each mode defines which components are required and in what configuration.
- Synedrix OS - Core Principles — Principle 11 (Open by default, sovereign by design) is implemented by the components in this stack.
Architecture
- Synedrix OS - Layer Taxonomy Table — The twelve canonical layers each platform maps to.
- Synedrix OS - Responsibility Matrix — Capability-by-capability ownership split.
Agent Specifications
- Synedrix OS - Agent Registry — The OS agents that run on this stack.
- Synedrix OS - Agent Runtime Policy Spec — Runtime policy schema and model routing rules.
Event Standards
- Synedrix OS - Event Taxonomy Standard — Event envelope and categories these platforms carry.
Synedrix OS · Tech Stack v1.8 · August 2026