Agents & the context engine

Connect your agent — three paths: Claude Code plugin · CLI · MCP. Each gives your agent access to durable graph-aware memory, live data (logs, traces, code, data sources) in KQL/SQL, and the graph that links them. See Coding agents overview for a comparison of all integration paths.

kyma also ships a built-in agent that turns natural-language questions into KQL/SQL over your live data — in the web app at /agent, the HTTP /v1/agent/ask endpoint, or the CLI.

Quick tour

# 1. Connect this terminal to a running Kyma server.
kyma connect http://localhost:8080 --token <bearer>

# 2. Ask a question — streams to stdout.
kyma query "what databases do we have, and how big is each one?"

# 3. Make Claude Code aware of your Kyma server.
kyma install-skill --also-link-claude
# Now in any Claude Code session, "ask Kyma what's slow today" works.

Three things to know

1. The provider is swappable. Pick Anthropic, OpenAI, Ollama, or "Claude Code CLI" (which inherits your local OAuth and tool loop) in /settings#engine. Each provider lists its installed models live — for Ollama, the picker calls ${host}/api/tags and shows what's actually on disk.

2. Skills are first-class. Drop a SKILL.md into one of the discovery roots and it shows up under /settings#skills, ready to enable. Enabled skills are injected into the system prompt for every agent turn (or, for the Claude CLI engine, picked up natively by Claude Code).

3. The CLI is the integration surface. kyma query "…" streams an answer to stdout. kyma install-skill writes a SKILL.md that teaches any coding agent to use this CLI as its window into your production data.

In this section

• Connect your agent — the three integration paths (plugin · CLI · MCP), the local single binary, and local↔server sync. Start here.
• Coding agents — comparison of all integration paths and when to use each one.
• Agentic Memory — persistent, graph-aware memory for agents: capture → LLM extraction with conflict resolution + bi-temporal validity → near-realtime hybrid recall. The memory_search tool, /v1/agent/memory/query, and live tuning.
• Engines — provider abstraction, model picker, credential resolution order, the /v1/agent/engine API.
• Claude Code engine — how the Claude CLI engine inherits Keychain OAuth, MCPs, and skills; when to use it.
• Claude Code plugin — hooks that auto-capture each session and inject recall into every prompt, plus /kyma-* slash commands.
• Skills — local discovery (./.skills/, ~/.kyma/skills/, ~/.skills/, plus opt-in Claude paths), the SKILL.md format, and how enabled skills are injected.
• Connect via CLI — install kyma, install the Kyma skill, and wire it into Claude Code / Cursor / Aider.

How does this relate to /v1/agent/ask?

/v1/agent/ask is the wire-level streaming HTTP endpoint. Everything above is built on top of it:

• The web /agent page is a React client of /v1/agent/ask.
• kyma query is a CLI client of /v1/agent/ask.
• The installed skill teaches other coding agents to shell out to kyma query, which is a client of /v1/agent/ask.

If you want to wire your own client, see Query → Agent endpoint for the SSE schema.