examples

Deuz SDK examples

Twenty-three runnable apps, each its own npm workspace. Start from the repo root:

npm install
npm run build            # examples import @deuz-sdk/core by name, so build dist first
npm run dev -w @deuz-examples/07-native-agent          # 07–23 run without an API key
export ANTHROPIC_API_KEY=sk-ant-…
npm run dev -w @deuz-examples/01-basic-stream          # 01–06 call Anthropic

Or use the convenience script: npm run example -- @deuz-examples/01-basic-stream.

Example	What it shows	API key
01-basic-stream	streamChat → textStream → await usage	ANTHROPIC_API_KEY
02-tool-loop	tools, maxSteps, a budget stop, and a self-healing tool error	ANTHROPIC_API_KEY
03-next-chat	Next.js App Router: toDeuzStreamResponse + useChat + an approval card	ANTHROPIC_API_KEY
04-structured-output	generateObject and streamObject from one zod schema	ANTHROPIC_API_KEY
05-durable-resume	a file SessionStore, a hard crash, resumeFromCheckpoint	ANTHROPIC_API_KEY
06-autonomous-agent	plan → delegate → run code → verify, with a live plan/activity feed	ANTHROPIC_API_KEY
07-native-agent	native runAgent: a validated output, a verifier, and an approval that suspends and resumes	none
08-dynamic-swarm	createSwarm with runtime spawn, a group blackboard, and createRounds	none
09-durable-ops	SQLite leases: recover() takes over a crashed executor's run, plus requestCancel and drain	none
10-evolve	evolve with SEARCH/REPLACE diffs and an evaluator cascade on SQLite, then a zero-call resumeEvolve	none
11-schedules-and-signals	createScheduler with a durable SQLite claim, and handleSignal with verifyHmacSignature	none
12-persistent-budgets	a per-user createSqliteBudgetStore scope shared by native runs and a swarm, until it denies	none
13-headless	@deuz-sdk/core/headless (2.3 preview): run / resume / inspect / doctor with JSON reports and exit codes, runs kept in SQLite; resume answers an approval, a call whose effect is unknown (--retry-call, --reconcile) and a failure in transit (--retry-failed)	none
14-native-memory-chat	a native run with memory (one recall, extraction once it completes), a chat projection and fallbackModels, then reconcileAgentRun repairing a failed chat write	none
15-native-ui-bridge	toDeuzAgentStreamResponse without a framework: readDeuzStream, agentStateFromParts and applyUIPart through an approval, a one-request resumeStreamAgent and the typed output, then a journal replay	none
16-goals-and-wakeups	@deuz-sdk/core/goal: a goal woken by a schedule occurrence and a webhook event, each claimed once in SQLite, with one leased native run per wake supplying the evidence	none
17-realtime-offline	@deuz-sdk/core/realtime on a scripted OpenAI Realtime socket: audio in, a tool call bridged to a Deuz tool, a spoken answer, usage priced per modality	none
18-batch-offline	@deuz-sdk/core/batch on a scripted Message Batches API: budget holds at submit, wait polling, per-result settlement at batch rates, an unpriced-model warning	none
19-mcp-server	createMcpServer: Deuz tools and an agent served over MCP 2026-07-28 to the Deuz MCP client in process, with an elicitation approving a gated tool	none
20-a2a-agent	a native run served over A2A with createA2AHandler + createA2AAgentExecutor: an in-process client answers its approval and gets the typed result as an artifact	none
21-versioned-memory-context	@deuz-sdk/core/memory/versioned and @deuz-sdk/core/context in a V3 native run (checkpointVersion: 3): asserted records that cite their sources, a replayed operation id, a revise and a stale-revision conflict, recordedAsOf, memory and working memory placed by the context compiler, a model write through the intent protocol, and a forget that redacts the stored context before the resume compiles again	none
22-goal-supervisor	@deuz-sdk/core/goal/runtime on one SQLite file: a durable schedule in Europe/Berlin caught up into the goal's inbox after a restart, the goal supervisor running a V3 run that waits for an approval (supervisor.resume), verifyGoalAcceptance in the verify hook, and the run's usage charged to the goal once	none
23-docker-runner	@deuz-sdk/core/compute/docker on the pinned alpine image: the sandbox's isolation defaults, a process recorded in SQLite with its output and an artifact, a reconnect that never starts it twice, cancel, reconcile and dispose; without Docker or the image it prints a skip message and exits 0	none (needs Docker)

Examples 01–06 call a real provider: each reads its API key from the environment at the app layer and passes it explicitly into the provider factory. Without the key, the Node examples exit with a message and the Next.js route answers 500 ANTHROPIC_API_KEY is not set. Core never reads process.env; that is deliberate, and it is why the same code runs unchanged on Node, Deno, Bun, and the edge.

Examples 07–23 need no key and no network, and print each step as it happens; 23-docker-runner needs a local Docker engine with its pinned image instead, and says so and exits 0 without them. Most run on createMockModel from @deuz-sdk/core/testing, a scripted model; 17-realtime-offline and 18-batch-offline script the provider's wire instead (a WebSocket through deps.webSocket, and the model factory's fetch), so the real session and batch code runs against a stand-in, and in 14-native-memory-chat the failing primary is an OpenAI model whose injected fetch answers 503 itself. A comment marked REAL PROVIDER in each index.ts of 07–12 and 14–22 shows the change that swaps in a real one (23-docker-runner calls no model); 13-headless takes the model from its command line instead: --model <id> runs that Anthropic model with the key from ANTHROPIC_API_KEY (a set key alone selects claude-opus-4-8), and its --mock scenarios run only when neither is given. 19-mcp-server depends on the v2 MCP SDK (@modelcontextprotocol/server and @modelcontextprotocol/client), which npm install adds with the workspace. The core test suite runs each of them from the sources and checks what its README says it shows: packages/core/test/examples-keyless.test.ts covers 07–12 and 14–20, examples-2-4.test.ts covers 21–23 (the full run of 23 only with DEUZ_TEST_DOCKER=1, its skip path always), and headless-example.test.ts covers 13-headless.

The Node examples (all but 03-next-chat, which runs next dev) run through Node's native TypeScript stripping (node --experimental-strip-types index.ts), so they need Node ≥ 22.6 and no build step or bundler of their own. Examples 09–13, 16, 22 and 23 also use node:sqlite, which ships unflagged from Node 22.13 and 23.4. Examples 09–12, 16, 22 and 23 keep their database in a temporary directory that each deletes when it finishes; 13-headless keeps its file between commands, so run and resume can be separate processes.