Run it, or embed it

Three steps to a live stack

git clone https://github.com/Joshwong1908/relionaut.git && cd relionaut
cat >> .env <<'EOF'
GLM_API_KEY=your-key            # any Anthropic-compatible endpoint
GLM_BASE_URL=https://open.bigmodel.cn/api/anthropic
GLM_MODEL=glm-4.6
EOF

docker compose up -d --build

Embedding the kernel

The kernel is a single directory — api/src/core/ — with zero npm dependencies and zero imports outside itself (enforced by a purity gate in CI). Vendor it into any Node 24 project (native TypeScript, no build step).

// 1. copy api/src/core → your project (e.g. ./agent-core)
// 2. implement the ports you need
const myTelemetry: TelemetrySource = {
  describe: () => ({ logsAvailable: false, metricsAvailable: true,
                     metricNames: ['checkout_error_rate'] }),
  queryMetrics: async (q) => /* call your time-series store */,
};

// 3. assemble and run
const agent = createAgent({
  llm: openaiProvider({ baseUrl, apiKey, model }),
  tools: [staticTools(...telemetryTools(myTelemetry))],
  store: inMemoryRunStore(),          // or your own RunStore
  system: { role, environment, dataSources, outputFormat },
  maxSteps: 12,
});
for await (const ev of agent.run({ runId, input, mode: 'read-only' })) { … }

Ports you don't implement simply don't exist: a telemetry source without queryLogs produces no query_logs tool — there is no "configured but unavailable" state, and the system prompt reflects reality automatically.

Governance & run modes

Every tool carries a risk tag (single source of truth) and a provenance string. A pure function decides each call inside the loop — the red line cannot be overridden by policy.

MCP tools without an explicit risk mapping are treated as write-risk — fail-closed. Map trusted tools via riskFor to make them available in read-only runs.

Audit trail

Every run ends in exactly one persisted terminal outcome. Exhausted the step budget? The store records step_budget_exhausted. The LLM endpoint failed mid-investigation? It records llm_error. Trajectories are stored as structured, block-paired turns — tool calls and their results replay in the exact shape the model saw.

// agent_messages row kinds (Postgres host)
text | tool_use | tool_result          // legacy-compatible audit columns
run_outcome                          // terminal: { outcome, summary, findings }
blocks JSONB                         // structured ChatMessage per turn

Environment variables

Security notes

Read-only by design: host commands pass a whitelist and a dangerous-character gate; the agent never mutates state and only suggests changes. Before exposing the platform to the public internet, read the auth-tunnel spec — token auth is mandatory, unauthenticated endpoints return 404.

Project resources