Splice
GitHub Get started

Composition

Splice skills are small tools with structured input and output. Composition is simply passing one tool's output into another tool's input. Splice provides no workflow engine, DSL or orchestration framework, and needs none: the agent (or your code) decides the order.

Skill A output → Skill B input

import { Splice } from "@spliceloom/sdk";

const splice = new Splice({ project: "./agent" });
await splice.init();
await splice.add("@splice/json");
await splice.add("@splice/http", { acceptPermissions: true });

// A: fetch raw text
const page = await splice.run("http.get", { url: "https://api.github.com/repos/spliceloom/splice-artifacts", responseType: "text" });
if (!page.ok) throw new Error(page.error.message);

// B: text → structured value
const parsed = await splice.run("json.parse", { text: page.output.text });
if (!parsed.ok) throw new Error(parsed.error.message);

// B: keep only what is needed
const picked = await splice.run("json.pick", { value: parsed.output.value, paths: ["full_name", "stargazers_count", "license.spdx_id"] });

Every step runs in its own sandbox process with only its own package's permissions: the JSON step cannot use the network even though the previous step could.

Runnable example

examples/agent-composition/compose.ts — deterministic by default (@splice/files → @splice/json), or --url for @splice/http → @splice/json. It prints each step and the final structured result; it needs no LLM or API key and is executed by the test suite against a local registry.

SPLICE_REGISTRY=https://registry.spliceloom.com node examples/agent-composition/compose.ts

Choosing tools from descriptors

A model-driven agent uses the descriptors, not hard-coded names:

const tools = await splice.tools();   // name, description, inputSchema, outputSchema, permissions
// give `tools` to the model; execute what it picks with splice.run(tool.qualifiedName, input)

outputSchema tells the model what the next step will receive, which is what makes chaining reliable. Over MCP the same happens automatically: an MCP client sees every installed tool with both schemas and calls them in sequence.

Guidelines

  • Keep intermediate data small: json.pick instead of passing whole documents around.
  • Prefer returning structured data from your own skills (declare output), so they compose.
  • Permissions do not add up across steps; each tool keeps its own. Grant the minimum per package.
  • Handle { ok: false } at every step; errors are data, not exceptions.