MVP / Alpha · v0.5.0 on npm

FabricFabricHarness

A headless TypeScript framework for durable, deployable autonomous agents. No TUI, no GUI — just an SDK and a CLI.

Get started →DatabricksUse casesGitHub
$npm install -g @fabric-harness/cli

Databricks · first-party

Build governed agents on Databricks

One databricks() call wires the agent's brain (Mosaic AI Model Serving), tools (governed SQL, Unity Catalog, AI Functions, Genie, Lakeflow, Vector Search, Feature Serving), durable state (Lakebase), and cost controls — all under a single Unity Catalog principal. Layered on top of UC; it never re-implements your permissions.

Mosaic AI servingUnity CatalogVector SearchGenieLakeflowLakebaseAgent Bricks
.fabricharness/agents/analytics-copilot.ts
1import { agent } from '@fabric-harness/sdk';
2import { databricks } from '@fabric-harness/databricks';
3

4// One call wires the brain, tools, governance, and state — under one UC principal.
5const dbx = databricks({
6  host: process.env.DATABRICKS_HOST!,
7  principal: { kind: 'service-principal', /* clientId, clientSecret, host */ },
8  model: 'databricks-meta-llama-3-3-70b-instruct',     // Mosaic AI Model Serving
9  warehouseId: process.env.DATABRICKS_WAREHOUSE_ID!,    // governed SQL
10  vectorSearch: { index: 'main.kb.docs', textColumn: 'chunk' }, // RAG
11  genie: { spaceId: 'space-123' },                      // NL → SQL analytics
12  governance: { stewardAudience: 'data-steward' },      // Unity Catalog + approvals
13});
14

15export default agent({
16  run: async ({ init }) => {
17    const fabric = await init({ modelProvider: dbx.modelProvider, tools: dbx.tools, policy: dbx.policy });
18    return (await fabric.session()).prompt('How many orders shipped last week?');
19  },
20});

Unity Catalog governance, RAG, deploy to Databricks Apps / Agent Bricks · full Databricks guide →

Databricks is first-party — Fabric also speaks every major model provider: Anthropic, OpenAI, Azure OpenAI, Bedrock, Google Vertex, Cohere, Mistral, Groq, and Cloudflare Workers AI. Model providers →

Authoring

Start fast. Go strict when you need to.

The bare @fabric-harness/sdk import gives you a working agent in 8 lines — headless defaults injected on every init(). Switch to @fabric-harness/sdk/strict for Temporal-backed durability or compliance workloads where every option must be declared.

.fabricharness/agents/echo.ts
1import { agent } from '@fabric-harness/sdk';
2

3export default agent<{ message: string }>({
4  name: 'echo',
5  triggers: { webhook: true },
6  run: async ({ init, input }) => {
7    const session = await (await init()).session();
8    return { reply: await session.prompt<string>(input.message) };
9  },
10});

Defaults injected on every init(): runtime: stateless, sandbox: virtual, loopRuntime: pi-agent-core, compaction: { enabled: true }. Override any value per call.

Sandboxes

Same agent code, swap the sandbox

One init({ sandbox }) call picks where the shell, filesystem, and tools execute. Edge, durable workflow, hosted, remote dev box, GPU — pick what fits.

.fabricharness/agents/cloudflare.ts
1import { agent } from '@fabric-harness/sdk';
2import { getCloudflareSandbox } from '@fabric-harness/cloudflare';
3

4export default agent({
5  run: async ({ init, env }) => {
6    const sandbox = await getCloudflareSandbox(env.SANDBOX, 'session-1');
7    const session = await (await init({ sandbox })).session();
8    return await session.prompt('hello edge');
9  },
10});

edge worker · full Cloudflare guide →

Use Cases

Built for the SDLC

Ten production-shaped agents, each with a runnable scaffold.

01

Issue Triage

02

PR Reviewer

03

Changelog Writer

04

Test Generator

05

Dependency Auditor

06

Schema Migration

07

API Docs Generator

08

Bug Reproducer

09

Release Notes

10

Incident Runbook

See all 10 →

Deploy

Author once, deploy anywhere

In-workspace on Databricks, on the edge, as a durable workflow, straight from your CI/CD pipeline, or any Node runtime — fabric-harness build --target emits the artifact, the runtime is pluggable.

Databricks

Run in-workspace as a Databricks App, or serve via MLflow for Agent Bricks / the Playground.

GitHub Actions

Run agents straight from your CI pipeline — triage, review, release notes on every push or PR.

GitLab CI/CD

Same agents in GitLab pipelines, gated by your existing CI rules and secrets.

Cloudflare Workers

Edge-deployed via @fabric-harness/cloudflare with R2 sources and KV/D1 stores.

Temporal

Long-running, retryable workflows that survive crashes and approval delays.

Node

Local dev, Docker images, AWS / GCP / Azure VMs, any Node 22+ runtime.

Docker

Sandboxed shell + filesystem isolation per agent. Useful for coding agents.

Render

One-command container deploy with a generated Blueprint and health checks.

Azure Foundry

Azure-managed hosted agent runtime, designed-for and rolling in.

Sources

Mount any knowledge base — no embeddings

Filesystem sources mount Markdown, MDX, or any text into a sandbox so the agent's built-in grep, glob, and read tools see it as ordinary files. Supports Fumadocs, Mintlify (local or hosted MCP), Cloudflare R2, S3, Azure Blob, and arbitrary HTTP.

.fabricharness/agents/runbook.ts
1import { agent, fumadocsSource, withFilesystemSources } from '@fabric-harness/sdk';
2

3export default agent<{ alert: string }>({
4  name: 'runbook',
5  triggers: { webhook: true },
6  run: async ({ init, input }) => {
7    const sandbox = withFilesystemSources('virtual', [
8      { mountAt: '/workspace/runbooks', source: fumadocsSource('./runbooks') },
9    ]);
10    const session = await (await init({ sandbox })).session();
11    return { plan: await session.prompt<string>(`Alert: "${input.alert}". Match a runbook.`) };
12  },
13});