Armor Wallet
Crypto, DeFi, Web3 infrastructure
An open-source MCP server and agent layer that lets AI assistants carry out real crypto operations: wallet management, swaps, DCA and limit orders, and staking.
AI Automation & Agents
Nanites
AI agents that investigate network incidents like a senior NOC engineer: parallel checks on real devices, correlated evidence with a confidence score, and a scoped fix for an engineer to approve.
- Root-cause accuracy on the incident evaluation set
- 94%
- A sub-agent per device, running read-only CLI checks
- Parallel
- Each finding carries evidence and a confidence score
- Explainable
- A scoped fix is proposed; an engineer applies it
- Safe fixes
The starting point
The challenge
Network incidents are slow to diagnose because the evidence is scattered across vendor devices, operating systems, telemetry, logs and configs. On-call engineers log into box after box to piece the story together, often in the middle of the night.
The solution
What we built
Built a ReAct-style investigation agent with LangGraph and LangChain. A Grafana alert triggers it, and it sets a goal and pulls the relevant context.
Spawns a sub-agent for each device, which connects and runs real read-only commands, so checks across devices happen in parallel.
Normalises vendor differences before comparing, for example IOS-XR layer-2 MTU against IP MTU, so a flagged mismatch is real rather than a reporting artefact.
Correlates the findings into an evidence list with a confidence score, then proposes a scoped fix for an engineer to review and apply.
Engineers can ask follow-up questions against the same investigation context.
Architecture
How the system works
- 01Grafana alertAn incident triggers the agent
- 02Agent managerSets the goal, gathers context
- 03Device sub-agentsParallel read-only CLI checks
- 04NormaliseVendor and OS differences
- 05CorrelateEvidence and confidence score
- 06Scoped fixAn engineer reviews and applies it
Engineering
How it was engineered
Built evaluation sets from recorded incidents and device outputs, including synthetic failure scenarios run on a network testbed.
Evaluated the agent's reasoning and root-cause findings after every iteration, locally and against the testbed.
Traced every run in LangSmith, from the incoming alert to the proposed fix.
Scope
What was delivered
01
Alert-triggered investigation agent
02
Per-device sub-agents for parallel checks
03
Vendor and OS normalisation logic
04
Evidence and confidence reporting
05
Follow-up questions on an investigation
06
LangSmith tracing
Gallery
Screens and diagrams
Technology
Stack
OpenAILangGraphLangChainLangSmithGrafanaPython
More work
Related case studies
AI assistant for a national diagnostics network
Healthcare, Diagnostics
An AI agent layer for a diagnostics network whose patient app has 1M+ downloads: plain-language lab-report explanations, test guidance, lab lookup and voice, orchestrated across MCP tool servers.
Intelligent service experience
Enterprise customer service
A governed AI service architecture in which Agentforce works inside Service Cloud. It is grounded in Data 360 context and Knowledge, scoped by topics and instructions, and always able to hand off to a human agent.
Need something similar?
We can scope and build a system with comparable capabilities, shaped around your own workflows, data and integrations.