Embeddable kernel
A single core/ directory — zero npm dependencies, zero upward imports, enforced by a purity gate. Vendor it, implement ports, assemble with createAgent().
Relionaut brings 2026-generation SRE agent capabilities — autonomous incident investigation, tool governance, bring-your-own LLM, MCP tools, persistent memory — in a portable kernel you drop into your own production systems. No platform lock-in.
Cloud SRE agents ship impressive capabilities bound to their platform. Open-source investigators bind you to their stack. Relionaut splits the difference: a self-contained agent kernel with the capabilities, and every integration point behind a port you own.
A single core/ directory — zero npm dependencies, zero upward imports, enforced by a purity gate. Vendor it, implement ports, assemble with createAgent().
Every tool declares risk; a pure govern() choke point blocks write-risk tools in read-only mode — structurally, not by prompt. Unknown MCP tools default to write (fail-closed).
Investigations distill into lessons; the next incident starts with the relevant history injected into context.
Any Anthropic- or OpenAI-compatible endpoint, Ollama for air-gapped. Your telemetry via one interface. Your toolchain via MCP.
Every run ends in a persisted terminal outcome — answered, step_budget_exhausted, or llm_error. Nothing is dropped from the audit trail.
Prefer batteries included? docker compose up starts the full self-hosted loop: collect → alert → investigate → auto-recover, with a web console.
The ReAct investigation loop, governance, memory, and audit live in the kernel. Everything environment-specific sits behind five interfaces — that's the whole integration surface.
The kernel — investigation loop, tool governance, prompt assembly, trajectory replay, memory injection.
LlmProviderstreaming chat + tool calls; provider protocols never leak past this portTelemetrySourcelogs & metrics queries; sources describe themselves into the promptToolSourcestandard tools, whitelisted shell, or any MCP serverRunStorebatched, block-paired trajectory persistence with terminal outcomesMemoryStorelesson record & retrievalThe first host — the self-hosted all-in-one is just one consumer of the kernel.
anthropic providerGLM / Claude-style endpointsPG telemetrybuilt-in collector: docker logs & stats, host metricsdocker toolsread-only whitelisted host forensicsPG run storeagent_messages with structured blocksalert enginerule evaluation → incidents → auto-investigate → auto-recoverFull stack: PostgreSQL, Redis, collector, alert engine, investigation agent, web console.
git clone https://github.com/Joshwong1908/relionaut.git
cd relionaut
echo 'GLM_API_KEY=your-key' >> .env
docker compose up -d --build
# open http://localhost:8082 — telemetry flows in 30s
Vendor the kernel directory, implement the ports you need, run an investigation.
import { createAgent, openaiProvider, staticTools, telemetryTools, inMemoryRunStore } from './agent-core/mod.ts'; const agent = createAgent({ llm: openaiProvider({ baseUrl: 'http://llm.internal:11434/v1', model: 'qwen3:32b' }), tools: [staticTools(...telemetryTools(myTelemetry))], store: inMemoryRunStore(), system: { /* role, environment, output format */ }, }); for await (const ev of agent.run({ runId, input })) { /* SSE events */ }
Open-source SRE agents cluster on an autonomy spectrum. Relionaut ships as a read-only investigator by default — the red line is structural — with the interface reserved for approved remediation as trust grows.
advise mode).approve-write + Approval).