Agent Platforms & Harnesses
The layer other agents on this site are often built on: platforms for building, orchestrating, and running multiple AI agents together, rather than a single-purpose product of their own.
I work at Fabrix.ai. Entries about Fabrix.ai products in this category aren't arm's length. See the full disclosure.
Fabrix.ai AgenticAI Platform
Fabrix.ai
Fabrix.ai's answer to why agent demos stall in production: a Context Engine caches large tool outputs and conversation history outside the LLM's window, so tools pass data to each other through a shared cache rather than through the model. A graph-based Ontology Layer maps how entities correlate across sources to drive what gets cached. A Tooling and Connectivity Engine generates MCP-compatible interfaces for systems without native MCP support. AgentOps adds cost tracking, observability, and human-in-the-loop approval.
Source of Truth
Agents need a reliable, structured record of what the network is supposed to look like before they can safely act on it. These are the inventory and intent-modeling platforms that play that role, plus the MCP servers that expose them to agents.
Nautobot MCP
kvncampos (community)
A community-built MCP server for Nautobot, the other major open source network source-of-truth platform, with both STDIO and HTTP deployment modes plus embedding search and RAG over Nautobot's data model. Unlike NetBox's official MCP server, this isn't built or endorsed by Nautobot's maintainer, Network to Code; treat it as a promising independent project rather than an official integration.
NetBox MCP Server
NetBox Labs (official)
NetBox Labs' own MCP server for NetBox, the widely used source-of-truth platform for network inventory, IPAM, and topology. Built with FastMCP, read-only by design, and works across every NetBox core object type (DCIM, IPAM, circuits, virtualization). Supports global search, field filtering to control token usage, and pagination, aimed at letting an agent query real network state instead of general knowledge.
Observability & Telemetry
Monitoring and metrics platforms, and the MCP servers or protocol bridges that let an agent query them directly instead of working from a stale dashboard export.
Hubble
Cilium (CNCF)
eBPF-based observability for cloud-native traffic. It isn't an AI agent, but it's increasingly the data source underneath higher-level agentic observability tools looking at Kubernetes networking.
LibreNMS MCP
mhajder (community)
A Python MCP server for LibreNMS covering devices, ports, alerts, inventory, locations, and logs, with both read and write operations rather than a read-only slice. Ships as a PyPI package and Docker image, with SSL verification, rate limiting, and logging controls built in. Smaller and newer than the Zabbix and Prometheus MCP servers here, worth watching rather than treating as mature.
Prometheus MCP Server
Prometheus (official org repo)
Lives under the official Prometheus GitHub org and lets an agent run PromQL queries, discover metrics, and analyze performance data, including against compatible backends like Thanos and Mimir. Optional TSDB admin tools sit behind an explicit flag, with a blunt README warning: it's not the server's fault if the LLM it's connected to deletes your data. A community alternative, pab1it0/prometheus-mcp-server, is actually more widely used, worth knowing both exist.
Zabbix MCP Server
initMAX (community)
An MCP server covering essentially the entire Zabbix API (223 tools across 58 API groups) rather than a narrow read-only slice: hosts, problems, triggers, templates, dashboards, and more. Supports multiple Zabbix servers at once, OAuth 2.1 and bearer auth, a read-only mode with per-token write protection, and systemd deployment, more production-hardened than most single-maintainer MCP servers in this space.
Orchestration & Execution
The layer where an agent's decision actually turns into a change on a device: task orchestration platforms and the MCP bridges that carry real write and execute capability, not just read access.
Junos MCP Server
Juniper Networks (official)
Juniper's own bridge between MCP clients like Claude Desktop or Copilot and real Junos devices, built on the junos-eznc SDK. Tools range from read-only (device facts, show commands) to a seriously privileged one: apply_config, which pushes configuration changes directly. Worth reading the permissions model closely before pointing it at anything you care about.
OPNsense MCP
vespo92 (community)
One of several OPNsense MCP servers, chosen here for how seriously it treats the risk of an agent misconfiguring a firewall: read-only by default, and write operations use OPNsense's savepoint mechanism, a 60-second countdown that auto-reverts the change unless explicitly confirmed. Also supports declarative plan/apply workflows for VLANs, firewall rules, DHCP, DNS blocking, and HAProxy.
Network Assurance & Digital Twins
Platforms that model network behavior end to end, so a proposed change can be verified against a simulated or mathematically modeled version of the network before it touches production.
Forward Networks
Forward Networks
Similar territory to IP Fabric: a digital twin used to verify proposed changes deterministically before they touch production. It's included for a different reason than most of what's covered on this page: a reminder that "verify before you act" doesn't always need an LLM in the loop, just good math and a reliable model of the network.
IP Fabric
IP Fabric
Discovers a network read-only over SSH and APIs and builds a mathematically accurate digital twin of its actual, current state. The agent-facing piece is an Enterprise MCP Server that lets tools like Claude or Copilot query that twin directly (end-to-end path lookups, shadow firewall rules, missing segmentation) instead of guessing from general networking knowledge.
AI Security
The threat models, hardening patterns, and access-control tooling specific to giving an agent real access to network infrastructure, distinct from traditional network security.
MCPSecBench
Academic research (arXiv)
A systematic security benchmark and testing playground for MCP, built around a taxonomy of 17 distinct attack types across four attack surfaces, evaluated against three major MCP hosts: Claude Desktop, OpenAI, and Cursor. Useful less as a product to adopt than as a concrete checklist of what "MCP security" actually needs to cover.
NoPorts
Atsign Foundation
Open source (BSD-3-Clause) remote-access tooling from the Atsign Foundation, carrying OpenSSF Best Practices and Scorecard badges plus SLSA 3 compliance, not just vendor claims. The noports-iosxe integration runs as a Docker app-hosting container directly on Cisco Catalyst 9000/8000 switches, giving SSH, gNMI, NETCONF, and RESTCONF access with no inbound listening port on the management plane.
OWASP MCP Top 10
OWASP Foundation
The first OWASP Top 10 dedicated to a single protocol surface, covering risks specific to MCP-enabled systems: model misbinding, context spoofing, prompt-state manipulation, insecure memory references, and covert channel abuse. Currently in Phase 3 beta and pilot testing, meaning the ten categories are stable enough to cite even though specific rankings and descriptions may still shift before a 1.0 release.
Standards & Protocols
The protocols and frameworks agentic networking is being built on, from the Model Context Protocol itself to early-stage IETF drafts and industry autonomy frameworks. Where something is a draft rather than an adopted standard, that's called out rather than implied.
Model Context Protocol
Agentic AI Foundation (Linux Foundation)
The open protocol underneath almost everything else on this page: a standardized way for LLM applications to connect to external tools and data over JSON-RPC. Originated at Anthropic in November 2024; as of December 2025, Anthropic donated it to the Agentic AI Foundation, a new Linux Foundation directed fund co-founded with Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare, and Bloomberg, so no single vendor now controls the spec.
NCFED (NetClaw-to-NetClaw Federation)
IETF (individual draft)
Lets independently operated AI network-engineering agents discover each other's capabilities and delegate tasks over authenticated sessions, including across frameworks (stated interop between OpenClaw- and Hermes-based agents). Documented as an individual IETF Internet-Draft, not a reviewed or adopted standard, but more detailed than that status implies: concrete wire framing, TLS 1.3 security, TOFU trust bootstrapping, and explicit handling of cross-agent prompt injection.
Network Management Agent (NMA)
IETF (individual draft)
A second, more institutionally backed IETF draft alongside NCFED: co-authored across CAICT, China Mobile, Huawei, Cisco, and ZTE, now on its fifth revision. Proposes a Network Management Agent that works alongside existing controllers rather than replacing them, aimed explicitly at reaching TM Forum's Level 4 autonomy, with four defined interface types for agent-to-user, agent-to-agent, agent-to-controller, and agent-to-network communication.
TM Forum Autonomous Networks Levels
TM Forum
The six-level (L0 to L5) autonomy framework referenced throughout this space whenever something claims to be "autonomous": L0 is fully manual, L3 is conditional autonomy for specific scenarios, L5 is full autonomy with no human involvement. Formalized across IG1230 (technical architecture) and IG1252 (levels evaluation methodology). Useful as a shared vocabulary and a reality check: almost everything on this page sits at L2 or L3, not L5, no matter what the marketing says.
Lab & Test Tooling
Sandboxes and lab environments built for trying an agent against network gear without risking anything real.
cml-mcp
xorrkaz (community)
An MCP server for Cisco Modeling Labs that lets an agent build lab topologies and query node status from plain-language requests. Handy for spinning up test environments without hand-writing topology files, and a good low-stakes way to try agent-driven networking without touching production.
NetworkOps_Platform
E-Conners-Lab (community)
A community project pairing NetBox with a large set of MCP tools spanning Cisco, Juniper, Nokia, Arista, and Linux devices, plus self-healing agents and drift detection against an intended state. A good snapshot of how far a single motivated maintainer can push multi-vendor agent tooling.
Community & Further Reading
Looser than the other categories by design: foundational reading for agent builders alongside the communities and venues where agentic networking is actively discussed.
Building Effective Agents
Anthropic
Not networking-specific, but foundational: Anthropic's own field notes on when to use a fixed workflow versus a genuinely autonomous agent, and the handful of composable patterns (prompt chaining, routing, orchestrator-workers, evaluator-optimizer) that cover most real use cases better than one large autonomous loop. Good background before evaluating any agent framework, including the networking-specific ones elsewhere on this page.
Network Automation Forum
Network Automation Forum
A practitioner-focused community for network automation and observability, running the AutoCon event series (AutoCon 0 in 2023 through AutoCon 6 planned for later in 2026) plus reports, webinars, and articles. Its community knowledge base, handyinfo, now includes an AI Agents learning-materials page linking back to this site.