MCP design choices
Token economics, wide schemas, and cell budgets behind Axiom's MCP server.
PLATFORM / MCP
Agent-native machine data observability.
Wire your AI agents to the same data your engineers query, with the same languages. Axiom's MCP server is designed around how agents actually use machine data: wide schemas, sensible cell budgets, and full access scope.
THE MOAT
Why design quality is the moat.
Every vendor ships an MCP server. The difference is how it's built — for the way agents actually consume machine data.
Metrics + agent queryability
High-cardinality metrics, fully queryable by AI agents through MCP and the metrics skill.
SRE skill
Patterns that move agents from hypothesis to proof.
Full access scope
No Standard Indexing SKU — every byte you load is queryable by your agents through the same primitives.
ARCHITECTURE / MCP SERVER
Wide schemas. Sensible cell budgets. Full access scope.
Datadog and ClickHouse already ship MCP servers; Splunk will. MCP existence is no longer a moat. The lasting differentiators are design quality (how the server formats results so agents don't burn tokens on real-world data), access scope (what data is reachable through the server), and managed delivery (whether the customer is running the cluster the agent queries).
TECHNICAL DEEP DIVES
Everything ships in the open protocol.
Six surfaces an agent-builder reaches for — each one a real APL or MPL primitive, not a per-vendor adapter.
Connect agents through the native MCP server.
Exposes APL (logs, events, traces) and MPL (metrics) to AI agents. Schema introspection, query execution, and result shaping designed for agent token budgets.
Move agents from hypothesis to proof with the SRE skill.
Patterns that move agents from hypothesis to proof during an incident: scope the symptom, find the surrounding signals, narrow the window, propose a root cause, verify against the data.
Give agents every byte you loaded.
No Standard Indexing SKU. Every byte loaded is reachable by the agent through the same primitives engineers use.
Run agent queries on the same usage-based dial.
Agent queries count toward query-hour usage. No separate rate limits, no second pricing dial. Pricing is the same usage-based model as APL.
“Axiom’s MCP won me round. During a pentest, Claude Code used it to investigate the logs and traces around our Sentry errors, then built a dashboard showing the requests, endpoints, and error codes the pentester was triggering.”
— Ed, Clove
FAQ
What agent-builders ask first.
DESIGN AND DIFFERENTIATION
Every major vendor has or will ship an MCP server. The differentiators are design quality (result shaping, cell budgets, schema introspection), access scope (no Standard Indexing SKU gating what agents see), and managed delivery (no cluster you operate for the agent to query).
The MCP server's result shaping is designed around agent token economics: cell budgets cap runaway responses, schema introspection exposes only what the agent needs to plan, and pagination is automatic for large result sets.
SKILLS
Patterns that move agents from hypothesis to proof: scope the symptom, find the surrounding signals, narrow the time window, propose a root cause, verify against the data. Used by SRE agents to lower MTTR during incidents.
Patterns for querying MetricsDB in MPL: aggregations, time-series, high-cardinality tag filtering, and comparison across time windows.
Skills are evolving as the MCP ecosystem matures. Today the SRE and metrics skills ship with the server. Additional skills (and a skill SDK) are on the roadmap.
INTEGRATION AND USAGE
No. Agent queries count toward query-hour usage the same way human queries do. The same in-app spending limits apply.
Yes. MCP is the open protocol; any MCP-compatible client (Claude, Cursor, custom agents built on the MCP SDK) connects.
No. Agents use the same query languages engineers use. The SRE and metrics skills teach the agent to compose queries that fit the question.
Built for engineers and their agents.
Petabyte-scale ingest. Pricing you can predict.SRE and metrics skills included.