skelm

Build secure, agentic, long-running workflows in TypeScript. Run them anywhere Node runs.

Why skelm exists

Most LLM workflow tools make security an afterthought. Agents call arbitrary tools, read arbitrary files, hit arbitrary URLs — because the framework has no model for preventing it. When something goes wrong (prompt injection, runaway loops, accidental secret exfiltration), you find out after the fact.

skelm is built the other way around. Every agent step starts with zero privileges. Filesystem roots, network hosts, MCP servers, CLI binaries, secrets — each is declared upfront in the step definition. Anything not declared is denied at dispatch, before the backend ever starts. The audit log records every privileged action in a tamper-evident chain so you can always reconstruct what happened.

The rest of the design follows from that principle:

• Real TypeScript. Workflows are .ts modules you type-check, refactor, test, and version like any other code. No DSL, no YAML, no JSON config.
• Three step kinds, none wrapping another. code() for deterministic logic, llm() for single inference calls, agent() for full multi-turn loops. Mix them freely in a single pipeline.
• Multi-backend agents. Opencode, ACP (Copilot, Claude Code, Gemini), OpenAI, Anthropic, Pi — plus a provider SPI for custom backends. Switch backends by changing one config key.
• First-class MCP support. MCP servers are lifecycle-managed by the gateway, not bolted on. Attach them per-step; the permission model governs which steps can reach which servers.
• Native control flow. parallel, forEach, branch, loop, wait, and nested pipelines are core primitives.
• Schedulable. Cron, interval, webhook, one-shot, and queue triggers. The gateway hosts everything long-running.
• Tamper-evident audit. Hash-chained audit log. Query it with skelm audit query.

Get started in 60 seconds

# 1. Install the CLI
npm install -g skelm

# 2. Scaffold a project
skelm init my-bot && cd my-bot && npm install

# 3. Run your first workflow
skelm run workflows/hello.workflow.ts --input '{"name":"world"}'

# 4. Stand up the gateway (long-running, handles scheduling + agent steps)
skelm gateway start

When this skill activates

Use this skill when:

• The user is working in a skelm project (any *.workflow.ts / *.pipeline.ts file)
• The user wants to scaffold, author, or modify a pipeline
• The user asks about AgentPermissions, skelm.config.ts, MCP wiring, backend setup
• The user wants to run, inspect, schedule, or debug a pipeline
• The user is migrating from another workflow tool (LangChain, Inngest, llm-task, lobster)

The unit of work

A pipeline is a TypeScript file that exports a pipeline() call:

import { code, llm, agent, pipeline } from 'skelm'
import { z } from 'zod'

export default pipeline({
  id: 'my-workflow',
  description: 'What this pipeline does.',
  input:  z.object({ task: z.string() }),
  output: z.object({ result: z.string() }),
  steps: [ /* Step[] */ ],
  finalize: (ctx) => ctx.steps['last-step'] as { result: string },
})

Step kinds: code · llm · agent · parallel · forEach · branch · loop · wait · pipelineStep · idempotent

Import everything from 'skelm'. Access prior step outputs via ctx.steps['step-id'].

Step kind quick reference

code() — deterministic logic

code({
  id: 'parse',
  run: (ctx) => ({ value: (ctx.input as { raw: string }).raw.trim() }),
})

llm() — single-shot inference

llm({
  id: 'classify',
  backend: 'openai',
  prompt: (ctx) => `Classify: ${(ctx.input as { text: string }).text}`,
  output: z.object({ label: z.string(), confidence: z.number() }),
  maxTokens: 512,
})

agent() — full agentic loop (default-deny)

agent({
  id: 'implement',
  backend: 'pi',
  prompt: (ctx) => `Implement ticket ${(ctx.input as { id: string }).id}. Return JSON {prUrl}.`,
  permissions: {
    allowedTools:       ['gh.*'],
    allowedExecutables: ['git'],
    allowedMcpServers:  ['github'],
    fsRead:             ['./'],
    fsWrite:            ['./src/'],
    networkEgress:      { allowHosts: ['api.github.com'] },
  },
  workspace: { mode: 'ephemeral', cleanup: 'on-run-end' },
  output: z.object({ prUrl: z.string() }),
  maxTurns: 20,
})

Default-deny: every AgentPermissions field defaults to deny when omitted. An agent with no permissions block cannot call tools, read files, execute binaries, attach MCP servers, or make network requests.

Permissions are part of the API

Composition is intersection-only. Project defaults → named profile → step-level. Each layer can only narrow, never widen.

Named profiles in skelm.config.ts:

defaults: {
  permissionProfiles: {
    'github-write': {
      allowedExecutables: ['git'],
      allowedTools:       ['gh.*'],
      allowedMcpServers:  ['github'],
      fsRead:             ['./'],
      fsWrite:            ['./'],
      networkEgress:      { allowHosts: ['api.github.com'] },
    },
  },
}

Apply: permissions: { profile: 'github-write', allowedTools: ['gh.create_pr'] }

Full permissions reference: {baseDir}/references/permissions.md

Project layout

my-project/
├── skelm.config.ts          # Required for gateway + agent steps
├── workflows/
│   └── hello.workflow.ts    # One pipeline per file
├── package.json             # { "dependencies": { "skelm": "^0.3.7", "zod": "^4" } }
└── tsconfig.json

Scaffold a new pipeline from template:

bash {baseDir}/scripts/new-pipeline.sh my-pipeline "What it does"
bash {baseDir}/scripts/new-pipeline.sh my-pipeline "What it does" --agent

Config reference: {baseDir}/references/config.md

CLI essentials

skelm run <workflow.ts> --input '<json>'   # run once
skelm list                                 # discover pipelines
skelm describe <id> --format mermaid       # visualize
skelm history --last 10                    # run history
skelm validate <workflow.ts>               # static preflight
skelm logs                                 # stream gateway logs
skelm audit query --run <id>               # tamper-evident audit trail
skelm schedule add <id> --cron '0 * * * *' # schedule
skelm gateway start                        # long-running gateway

Exit codes: 0 ok · 1 CLI error · 2 schema validation · 3 run failed · 4 cancelled · 5 wait timeout · 6 permission denied · 7 step timeout

Full CLI reference: {baseDir}/references/cli.md

The gateway is the trust boundary

The gateway owns permission resolution, enforcement, secret resolution, audit log, approval gating, trigger dispatch, and registry management.

Never write permission enforcement in pipeline or step code. Pipelines are the user layer. The gateway is the trust layer.

skelm gateway start
skelm gateway status
skelm gateway install --systemd   # systemd unit at ~/.config/systemd/user/skelm-gateway.service

Gateway reference: {baseDir}/references/gateway.md

Common pitfalls

• Widening at step level — networkEgress: 'allow' in a step when the project default is deny has no effect. Intersection always wins.
• Missing Zod schema — input/output are validated at run boundaries; omitting them skips validation silently.
• agent() with an unregistered backend — step fails at runtime if backend references an id with no matching entry in config backends: or instances:. The pi SDK backend must be in instances:.
• Step id collisions inside parallel() — sibling ids must be unique within the parallel block.
• Editing dist/ — never edit generated files. Run pnpm build to regenerate.

Full references

• {baseDir}/references/pipeline-authoring.md — all builders, control flow, context shape, retry
• {baseDir}/references/agent-step.md — agent() signature, backends, workspace modes, MCP
• {baseDir}/references/permissions.md — full permission model, TrustEnforcer, testing
• {baseDir}/references/config.md — skelm.config.ts shape, backends, MCP entries
• {baseDir}/references/gateway.md — gateway lifecycle, HTTP surface, audit log, systemd
• {baseDir}/references/cli.md — complete CLI reference with all flags and exit codes

skelm v0.3.7 · MIT · npm