KuraPre-release

Usage

Chat over HTTP

One-shot query:

curl -s http://127.0.0.1:19192/v1/chat/query \
  -H 'content-type: application/json' \
  -d '{
    "query": "summarize my day",
    "provider": "claude_code_cli",
    "threadId": "thr_personal"
  }'

Streaming (SSE):

curl -N http://127.0.0.1:19192/v1/chat/query/stream \
  -H 'content-type: application/json' \
  -d '{"query": "hello", "provider": "echo"}'

Every turn runs the full pipeline: skills + persona compile into the prompt, thread continuity injects prior turns, the context plugin injects memory bootstrap and query-time recall, the session-strategy plugin shapes the window under budget — and whatever the model sees is exactly what is persisted on the dispatch record (model-visible = logged).

Tools

A turn can call tools. What the model is offered is resolved per turn and per tenant:

  • MCP tools — everything the servers you connected publish, authorized under the chat surface’s exposure rules; a tool marked approval_required waits for a person to answer instead of failing.
  • memory.lookup — recall over your own Ready memory, the same fused ranking as POST /v1/retrieval/queries; hits cite Memory[<layer> <id>].
  • web.search / image.generate / video.generate — one tool per configured tool profile. Kura does not pick a vendor: the mcp_backed family forwards to an MCP server you chose (the model sees that tool’s real argument schema), and every call runs through the quota and egress gates. builtin_stub proves the path without a vendor account.
curl -s http://127.0.0.1:19192/v1/tools/profiles -H 'content-type: application/json' -d '{
  "title": "my search", "capability": "web.search", "family": "mcp_backed",
  "authMode": "none", "mcpServerId": "mcp_search", "mcpToolName": "search", "isDefault": true
}'

Every call is a chat.tool.called event and a kura_chat_tool_calls_total metric, and the chat/tool-call hook can rewrite or veto it.

Browser

computer-use sessions drive a real browser when the plugin is configured with driver: "subprocess": a supervised Node worker (capabilities/browser/worker.mjs, Playwright + Chromium) speaks a line-JSON protocol to the daemon; hangs and crashes are reported to the capability supervisor and the worker is respawned.

TypeScript SDK

import { createKuraClient } from "@kura/client";

const client = createKuraClient({ baseURL: "http://127.0.0.1:19192" });

const result = await client.queryChat({ query: "hello", provider: "echo" });

// Streaming
for await (const chunk of client.streamChatQuery({ query: "hi" })) {
  process.stdout.write(chunk.delta);
}

// Memory, plugins, retrieval…
const plugins = await client.listPlugins();
const recall  = await client.queryRetrieval({ query: "package manager" });
const assets  = await client.listMemoryAssets({ layer: "l1", status: "ready" });

Terminal UI

kura tui launches the full-screen terminal client: conversations, thread continuity, and a live daemon event stream (/events).

Web operator shell

kura web serves the installed web shell locally and opens your browser. It is the operator console: memory assets + review queue, Plugins (assembly report, enable/disable, hooks), channels, routines, providers, quota, diagnostics, evaluation, and support surfaces.

Threads and continuity

Pass a threadId to keep continuity: prior turns are stored as governed continuity records and re-rendered into each dispatch (bounded, redaction safe). IM channels get one context per thread automatically.

Events

Everything observable is an event (llm.*, memory.*, context.*, chat.*, improvement.*, system.*), persisted in the store ledger and published on the live bus. The TUI /events view and the sessions API expose them.

Last updated: