Observed Signal · Jul 16, 2026 · Technical Release · Source: DEV Community · Impact: 2/5 · Sentiment: Positive
Costory rebuilds FinOps MCP for LLM-friendly queries
Costory describes lessons learned while building a FinOps Model Context Protocol (MCP) server to let large language models (LLMs) query normalized cloud cost and metric data. The first MCP mirrored internal APIs (multiple endpoint tools and nested JSON filters), which caused the LLM to misuse endpoints, produce syntax errors, and repeat questions. Costory replaced many specific endpoints with a single composable query tool, moved from nested JSON filters to CEL filter strings, added a get_context first-call to reduce repeated prompts, and rewrote error messages to be human- and model-friendly. These changes reduced model error rates (filter errors down ~25%), improved UX, and produced a new app feature (Advanced Explorer). The MCP integrates with BigQuery and is compatible with MCP-capable clients (Claude, ChatGPT, Cursor).
Describes practical engineering patterns for LLM-facing APIs and FinOps tooling (single composable query tool, CEL filters, context priming, clearer errors). Useful to engineering teams integrating LLMs, but not a major platform policy or industry-shifting announcement.
Track Google Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Costory ingests billing data from multiple cost centers including AWS, GCP, Azure, Cursor, Anthropic, Aiven, and Kubernetes Cost and normalizes it using the FinOps Foundation FOCUS format.
- The initial MCP exposed six tools mirroring internal endpoints: query_cost, query_cost_diff, search_filters, get_metric, query_metric, and create_alert.
- Costory replaced multiple endpoint-specific tools with a single composable query tool supporting query types like cost, metric, usage, externalMetric, budget, and formula.
- They replaced nested react-querybuilder JSON filters with a single CEL filter string (filterCel), reducing model error rate on filters by roughly 25%.
- They implemented a get_context first-call and rewrote error messages to be model-friendly; runtime validation and 'did you mean' suggestions are sourced from BigQuery.
Connected Companies & Entities
7 Entities mapped“our runtime already runs on BigQuery, so the valid values and the "did you mean" suggestion were just a passthrough from BQ errors....”
“Costory pulls in billing data from all your cost centers (AWS, GCP, Azure, Cursor, Anthropic, Aiven, Kubernetes Cost), normalizes it into on...”
“Costory pulls in billing data from all your cost centers (AWS, GCP, Azure, Cursor, Anthropic, Aiven, Kubernetes Cost), normalizes it into on...”
“Costory pulls in billing data from all your cost centers (AWS, GCP, Azure, Cursor, Anthropic, Aiven, Kubernetes Cost), normalizes it into on...”
“We looked at how other dashboarding solutions handle this: we saw Grafana MCP had similar issues and followed the same pattern: a diff of op...”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Model Context Protocol (MCP) Fundamentals Guide
This technical tutorial introduces the Model Context Protocol (MCP), an open standard for connecting large language models (LLMs) to external tools, data sources and services. It demonstrates building an 'Analyzer' MCP server using the FastMCP framework, explains MCP message types (tool discovery and tool execution), and shows transports (stdio, HTTP, WebSockets). The post describes moving from local development to production via Bedrock AgentCore Runtime—containerizing MCP servers, registering them with a runtime client, and securing access with Amazon Cognito. It also shows how Strands Agents can consume remote MCP tools as if local, and outlines best practices: descriptive docstrings, strict Python type hints, error handling, logging, and composable tool design to enable chaining and context awareness. The article targets developers building reusable, secure, scalable agent-accessible tools across multiple LLMs (e.g., Claude, GPT, Nova).
MCP Reframes How LLM Agents Use Tools
The article argues that the Model Context Protocol (MCP), introduced by Anthropic, is changing how large language model (LLM) agents integrate with external services by replacing the REST-centered integration mental model rather than HTTP itself. MCP is described as a session-oriented, bidirectional protocol that enables capability discovery, persistent sessions, server notifications, and multi-step stateful interactions without bespoke client glue code. The author highlights real-world adoption pressure (including a Cognizant–Anthropic expansion), urges builders to test the MCP TypeScript SDK, and warns that server implementations vary in quality, recommending defensive client-side handling.
Model Context Protocol Eliminates Integration Glue Code
This technical deep dive (Part 1 of 15) introduces the Model Context Protocol (MCP), a small JSON-RPC protocol designed to replace bespoke agent-to-backend integration glue with a discoverable, capability-first model. Using a running example called Mattrx (a multi-tenant marketing-analytics SaaS), the author shows that MCP turns N×M bespoke integrations into N+M servers, centralizes auth/audit with a single OAuth/Entra identity boundary, enables runtime tool discovery, and creates a safe, scoped path for external AI assistants. Reported benefits in the running system include collapsing 14 point-to-point integrations into 3 MCP servers, deleting ~9,000 lines of glue code, reducing onboarding from ~3 days to ~2 hours, and cutting agent tool-call errors from 6% to 0.8%. The protocol surface is intentionally small (initialize, tools/list, tools/call) and supports multiple transports (stdio for local dev; streamable HTTP + SSE in production).
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.