Local web workspace
The local web workspace is reworked around review: a sessions drawer, light and dark themes, a movable diff panel, and a searchable model selector. zhivex-code project add registers a folder for a running local server.
Code 0.4 · Harness 1.5 · Open source (MIT)Delegate the code
Delegate the code
Keep control
A coding agent governed by design. It works inside a sealed boundary, and nothing reaches your repo until you approve the exact patch.
❯ zhivex-code
zhivex code 0.4 · harness 1.5
task › Fix the failing tests in src/billing
✓ context bound repo acme/api @ 4f2c9e1 · src/billing/**
✓ sandbox: opt-in OCI snapshot · no network · no secrets
› running tests… 2 failing invoice.test.ts
? approval run `pnpm test src/billing` in sandbox? yes
--- a/src/billing/invoice.ts
+++ b/src/billing/invoice.ts
@@ -41,7 +41,7 @@
- const total = items.reduce((s, i) => s + i.price, 0);
+ const total = items.reduce((s, i) => s + i.price * i.qty, 0);
- return round(total * tax);
+ return round(total * (1 + tax));
✓ tests pass 14 passed · 0 failed
? Import this exact patch? ✓ approved
✓ patch admitted sha256:9c1e…a7 · no git writes made
How it works
Four steps · One boundary
Your approval at the gate
Every task follows the same governed path, from context to proof.
01
Bind context
The task binds to a repo, a commit and the files in scope. No drifting context.
02
Work in the boundary
The default runtime exposes no host shell. Isolated commands run in an opt-in OCI snapshot with no network and no secrets.
03
Approve
Sensitive steps pause for you. You see the diff and test results before anything moves.
04
Admit & prove
The exact approved patch is admitted, with a record of what ran and what changed.
Boundaries
What the agent can't do
Limits are enforced by the runtime, not requested in a prompt.
$ sh -c "…" on hostNo arbitrary host shell
The default runtime exposes no shell. Isolated commands stay in the configured sandbox.
curl https://…No container network
With isolated execution on, the container has no network.
git push origin mainNo automatic Git writes
Commits and pushes stay yours. The agent proposes a patch.
apply --force stale.patchNo silent stale-patch overwrite
If your tree moved, the patch is rejected, not forced.
17 OCTOBER 2026 · Harness 1.5.0
What's new in 1.5
Harness 1.5.0 and Code 0.4.0 are on npm latest. The Code CLI is zhivex-code. Published releases are verified with npm provenance. These notes follow the repository changelogs. Bedrock and Azure OpenAI are provisional and not live-certified.
Task and evidence
The local web client adds a task and evidence panel: goal, revision, artifact and check provenance, task-wide accounting, and explicit accept, correct, and continue. Web task control negotiates task.control.v1.
Audit pack
zhivex-code export-audit and verify-audit, and the same zhx commands, use the experimental @zhivex-ai/harness/audit entrypoint. Verification checks the pack without resolving provider, store, or budget configuration.
Bedrock and Azure OpenAI
Provisional routes on the matched engine. Bedrock Converse reports Tier B. BEDROCK_RUNTIME=openai and Azure OpenAI are the Tier A paths. These routes have no live certification.
Security hardening
restore_file is pinned to a destination and digest. Each budget ledger stays bound to its source namespace. Protected paths compare in locale-independent lowercase. High-entropy output from credential-bearing MCP servers is flagged.
Release provenance
Published Harness and Code releases are verified with npm provenance. npm latest for this pair is Harness 1.5.0 and Code 0.4.0.
Build on the Harness · @zhivex-ai/sdk 1.34.0
The integration layer
underneath Zhivex Code
Harness’s integration layer, and a standalone TypeScript SDK. The same generateText, streamText, tools, and durable agents run on Bun and Node.js. Install only the adapters you use. 1.34.0 adds an experimental MCP server.
14providers
274models
one npm install@zhivex-ai/sdk
Default model catalog in Zhivex/zhivex-ai-sdk@c3f4bf58d2bd1fee8fc0d3be44b432507350b95c, packages/sdk/src/catalog/fragments.ts. 14 provider fragments, 274 entries.
$ npm install @zhivex-ai/sdk @zhivex-ai/openaiMIT · SLSA provenance · 1.34.0
The Harness runs the governed coding agent. This SDK is that integration layer and a standalone SDK: generation, tools, durable agents, an experimental MCP server, and @zhivex-ai/sandbox-docker 0.2.0.
import { streamText } from "@zhivex-ai/sdk"; const result = streamText({ provider: "openai", prompt: "Say hello from the unified SDK." }); for await (const chunk of result.textStream) { process.stdout.write(chunk); }
Illustrative stream. The provider id is the only line that changes.
14 provider adapters · One contract
OpenAITier A
Azure OpenAITier A
BedrockTier B / A by runtime
xAITier B
MetaTier B
AnthropicTier B
GeminiTier B
VertexTier B
QwenTier B
DeepSeekTier B
Z.aiTier B
OpenRouterTier C
KimiTier C
OllamaTier C
Tier A approval-capable hosted tools or remote MCPTier B strong tool-using agentsTier C basic tool loops
@zhivex-ai/sdk@zhivex-ai/core@zhivex-ai/agents@zhivex-ai/react@zhivex-ai/gateway@zhivex-ai/openai@zhivex-ai/xai@zhivex-ai/meta@zhivex-ai/azure-openai@zhivex-ai/anthropic@zhivex-ai/gemini@zhivex-ai/vertex@zhivex-ai/qwen@zhivex-ai/kimi@zhivex-ai/deepseek@zhivex-ai/zai@zhivex-ai/openrouter@zhivex-ai/bedrock@zhivex-ai/ollama@zhivex-ai/sandbox-docker
generateText · streamTextPortable generation
The same calls on every adapter. Switching providers means changing the model line.
tool() · toolChoiceTools
Callable tools with tool choice across adapters.
generateObject · streamObjectStructured output
Schema-validated objects, native where the provider supports it.
reasoning: { effort }Reasoning
One effort ladder, mapped to each provider’s native controls.
Hosted tools / MCPHosted tools
Provider-hosted tools, web search and remote MCP where available.
Files API · multimodalFiles & media
Files, image, audio and realtime sessions on supported adapters.
resumeAgent · sessionsDurable agents
Sessions stored in Postgres, SQLite or a file you operate.
getApiStability()Stability contract
Every export is stable, beta or experimental. CI fails if one is unclassified.
@zhivex-ai/gatewayGateway
In-process policy routing and fallback across registered adapters.
createAgentApprovalQueueApprovals (beta)
Human approval queues for agent actions.
@zhivex-ai/reactReact
Headless chat state, SSE transport and accessible components.
Bun · Node.jsServer runtimes
Node.js, Bun, route handlers or workers. Keys stay on the server.
@zhivex-ai/sdk/mcp-serverMCP server
Experimental. Local tools and agents over Streamable HTTP and newline-delimited stdio. Tools that require approval are not executed.
@zhivex-ai/sandbox-dockerDocker sandbox
Experimental reference environment, version 0.2.0. Each run gets a non-root container, network off by default, resource limits, snapshots, and lease-aware resume.
import { createEgressPolicy, generateText } from "@zhivex-ai/sdk";import { createOpenAI } from "@zhivex-ai/openai"; const apiKey = process.env.OPENAI_API_KEY;if (!apiKey) throw new Error("Set OPENAI_API_KEY in the server environment."); const egressPolicy = createEgressPolicy({ allow: [{ provider: "openai", model: "gpt-6-astra", region: "*", endpoint: "api.openai.com" }]}); const result = await generateText({ model: createOpenAI({ apiKey })("gpt-6-astra"), prompt: "Describe Zhivex AI SDK in one sentence.", egressPolicy, egress: { endpoint: "api.openai.com" }}); console.log(result.text);
import { generateText } from "@zhivex-ai/sdk";import { createAnthropic } from "@zhivex-ai/anthropic";import { createOpenAI } from "@zhivex-ai/openai"; const openai = createOpenAI({ apiKey: process.env.OPENAI_API_KEY });const anthropic = createAnthropic({ apiKey: process.env.ANTHROPIC_API_KEY }); const prompt = "Respond in one short sentence."; const fromOpenAI = await generateText({ model: openai("gpt-4o-mini"), prompt });const fromAnthropic = await generateText({ model: anthropic("claude-sonnet-5"), prompt }); console.log(fromOpenAI.text);console.log(fromAnthropic.text);
import { generateText } from "@zhivex-ai/sdk";import { createOpenAI } from "@zhivex-ai/openai"; const result = await generateText({ model: createOpenAI({ apiKey: process.env.OPENAI_API_KEY })("gpt-6-astra"), prompt: "Review this architecture and identify its highest-risk assumption.", reasoning: { effort: "high" }});
Source: Zhivex/zhivex-ai-sdk · README.md · Quick Start
Design partners
Shape a governed coding agent with us
We're opening 10 design partner spots for teams who want to delegate real work without giving up control.
10 spots