Armature: Product Analytics & Evals for MCP Agent Sessions

Armature provides deep product analytics specifically designed for agentic sessions. It allows you to observe exactly how users interact with your product through interfaces like Claude, ensuring you aren't flying blind when your users delegate tasks to an AI.

"User sessions on your Claude Connector, ChatGPT App, or MCP server occur within the AI's client—not your own UI. Armature bridges this gap by capturing user intent, agent reasoning, and execution paths."

🔄 The Evolution of Visibility

Previously, developers relied on fragmented logs. Armature transforms those logs into a coherent narrative.

~~Scattered tool call logs~~ $\rightarrow$ Reconstructed Sessions

Comparison: The Experience Shift

Feature	Without Armature	With Armature
Visibility	Raw, disconnected logs	Full session reconstruction
Intent	Guessed from API calls	Explicitly captured
Reasoning	Invisible "black box"	Agent thinking logs visible
Outcome	Manual log auditing	Automated success scoring

🚀 Key Capabilities

1. Discover Top Use Cases

Armature utilizes specialized models to analyze every session and categorize the user's primary objective. These are then ranked by frequency and success rate.

Supported: $\text{Success Rate} = \frac{\text{Completed Tasks}}{\text{Total Attempts}}$
Unsupported: Armature identifies requests users tried to make that your agent couldn't handle.

Example Use Case Distribution:

Create & Send Invoices: 38%
Payment Reconciliation: 22%
Bulk Refunds: 14% (Currently not supported)
Revenue Export: 9%

2. Proactive Issue Identification

Don't be fooled by a 200 OK response. Armature scans for "silent failures," such as logic loops or dead ends, grouping them by root cause.

Critical Issues:
- 127 $\rightarrow$ Agent loops due to missing auth scopes $\uparrow$ 43 this week $\text{ [Fix it] } \rightarrow$
- 64 $\rightarrow$ Search fails on "refund" terminology
- 31 $\rightarrow$ Pagination causing truncated exports
- 12 $\rightarrow$ Rate limits on bulk updates

3. High-Fidelity Session Replay

Every interaction is assigned a quality score (e.g., a 92 score). You can dive into any session to see the full trace:

Trace Example:

User Intent: "Set up a paid plan for ACME and invoice them"
Action: list_customers $\rightarrow$ get_plans
Agent Thinking: “Invoice needs a billing contact — retrying with the owner email.”
Action: send_invoice $\rightarrow$ 402 Error $\rightarrow$ send_invoice (Retry)
Result: $\checkmark$ Task completed successfully.

🛠️ Integration Guide

Getting started takes only a few minutes. Follow these steps:

Step 1: Generate Credentials Create your account and obtain your unique API key. ARMATURE_API_KEY = "arm_9f3k••••••••"
Step 2: Configure Agent Provide your coding agent with a prompt to integrate the SDK into your deployment secrets.
Step 3: Deploy & Monitor Push your changes and watch sessions flow in.

# Example: Adding the key to your environment
export ARMATURE_API_KEY=arm_9f3kxxxxxxxxxxxx
npm run deploy

🛡️ Security & Privacy

We treat all session data as highly sensitive. To protect your users, Armature employs detection models that redact PII (Personally Identifiable Information) and secrets by default before the data is ever stored.

Redaction Example:

Original: "Invoice ACME and send it to jane@acme.com using key sk_live_9f3kq2…"
Processed: "Invoice ACME and send it to [PII] using key [SECRET]"

💳 Pricing Plans

Feature	Free	Custom
Monthly Cost	`$0`	Contact Us
Included Sessions	1,000 / mo (then $\$ 50/1\text{k}$)	Tailored
Projects/Users	Unlimited (Launch Offer)	Unlimited
Data Retention	7 Days	Custom
Core Tools	Dashboard, Replay, Grouping	Everything in Free
Enterprise	$\times$	SSO/SAML, SLA, Audit Logs, Dedicated Onboarding

❓ Frequently Asked Questions

Q: How does Armature actually capture the data? A: You integrate our SDK into your backend (MCP server, ChatGPT App, or Claude Connector). The SDK wraps your existing logic to capture sessions without altering the user experience. It works across any client that can reach your server, including:

Claude
ChatGPT
Cursor
Codex
Gemini CLI

Q: How is this different from tools like PostHog, Amplitude, or Mixpanel? A: Traditional analytics track what happens in your UI. However, when a user uses an agent, the interaction happens inside the AI client. Armature is built specifically for the engineers building these agents, showing how the agent experiences your product.

Ready to optimize your agent experience? [Start for free $\rightarrow$ ] | [Book a demo]