Workflow MCP

workflow-mcp is a standalone runtime and MCP server that executes Claude Code dynamic workflow files — the JavaScript orchestration programs that fan out dozens or hundreds of agent calls and return a single result — without requiring Claude to run them.

The whole project hangs on one promise:

A workflow authored for workflow-mcp can be copied into .claude/workflows/<name>.js and run by a compatible Claude Code release without changing the file — and a real Claude workflow file runs through this runtime without any importer or translation step.

Everything that isn't portable — which provider executes the agents, its credentials, the durable run cache, MCP run IDs, and UI state — lives in the runtime, never inside the .js file.

Why it exists

Claude Code workflows are a strong primitive: the JavaScript owns the loops, branches, fan-out, and aggregation, while each agent() call owns the model reasoning and side effects — keeping hundreds of intermediate results out of the parent conversation. But out of the box they come with three constraints:

• Claude executes every agent node. There is no seam for another provider.
• A run belongs to a Claude session. Its state lives inside that session directory, and it does not survive as an independently controllable object.
• There is no server surface. Other tools cannot discover, launch, follow, or resume a run.

workflow-mcp keeps the exact same workflow file and lifts those constraints: the same .js runs through a provider-neutral engine (Codex today), every run is a durable object that survives restarts, and any MCP-capable host can drive it over a stable set of tools.

How it works

Two kinds of portability are kept deliberately separate — MCP solves the first, the provider interface solves the second:

MCP-capable host
      │  workflow_list · workflow_run · workflow_run_events · …
      ▼
WorkflowService  +  durable event store   ← the long-lived owner of runs
      ▼
Claude-compatible JavaScript runtime       ← agent/parallel/pipeline/phase/…
      ▼
provider-neutral AgentProvider
      ├─ Codex SDK  (@openai/codex-sdk)
      ├─ fake       (deterministic tests)
      └─ future providers

• Claude-compatible runtime. A restricted, killable Node/V8 context that exposes exactly the globals Claude injects (agent, parallel, pipeline, phase, log, workflow, args, budget, top-level await/return) with the same discovery rules, metadata grammar, and cache identity. The behaviour is pinned to an observed Claude Code profile so future Claude releases add a new profile instead of silently breaking old runs.
• Durable service. The WorkflowService — not any single MCP connection — owns runs. Every event is appended and fsynced before any subscriber sees it, so a run can be reconstructed after a renderer reload, a provider reconnect, or a process restart by replaying a strict event cursor.
• Provider-neutral execution. The engine only knows an AgentProvider interface. The first real adapter drives the official Codex SDK; a deterministic fake provider runs the conformance suite. Model aliases (haiku/sonnet/opus) are host policy, never guessed.
• Reliability. One work-conserving scheduler across all runs, supervised per-agent retries, one process-owned Codex host per attempt, a shared provider circuit breaker, single-writer fencing, and interrupted-run recovery that sparsely reuses already-completed siblings.

What you get

• Portable workflow files — the same .js runs here and in Claude Code.
• A durable MCP server — fourteen stable tools (workflow_list, workflow_author_guide, workflow_describe, workflow_validate, workflow_run, workflow_run_status, workflow_run_events, workflow_result_read, workflow_run_cancel, workflow_resume, workflow_agent_list, workflow_agent_result_read, workflow_agent_results_read, workflow_agent_transcript_read) over stdio or an authenticated loopback HTTP transport.
• Per-agent inspection — a finished run is not just its final value. List its logical agents with attempt history, then read any single agent's complete untruncated output, or sweep them all in one paginated walk.
• Immediate run handles — workflow_run returns a run ID at once; clients follow progress by polling a durable cursor, not a transport-specific push.
• Unattended best-effort completion — retryable read-only work restarts in a fresh provider thread. An exhausted or unsafe logical assignment becomes a versioned __workflowAgentFailure coverage gap, while independent siblings and final synthesis continue. Such runs finish as completed_with_errors; only persistence or supervisor faults fail the complete run.
• Resume — continue a managed run, or import-and-resume a real Claude run after verifying its source and journal byte-identity. Exact source/arguments reuse completed calls sparsely; automatic crash recovery also preserves terminal coverage gaps, while an explicit manual resume retries those gaps. Edited source retains the longest unchanged prefix. MCP callers may pass a managed run_* ID or Claude's native wf_* ID; Claude's own files are never rewritten. For exact-source Claude imports, bounded hashes of the original subagent prompts preserve completed dynamic-pipeline siblings even when cached parents settle in a different order; raw prompt text is not copied into the workflow-mcp sidecar.
• An embeddable service — the same WorkflowService and tool registrar that the CLI uses can be mounted inside another host (this is how Agent Code renders each run as a live feed card) instead of starting a second server.
• A browser-safe state entry — workflow-mcp/state exposes the event union and pure reducer with no filesystem, MCP, or Codex code, so a renderer can project run state without pulling server code into its bundle.

Getting started

The Docker-first standalone product needs no host Node or Codex installation. It ships one project-scoped daemon, a Codex MCP proxy, terminal UI, optional local web UI, isolated credentials, and durable named-volume state. Start with the verified release and operator guide in standalone/README.md; the full decisions and support boundaries are in standalone/docs/adr.

For core-library development, Node ≥ 20.19 is required. A source checkout can build and test both the provider-neutral runtime and isolated standalone package:

npm install --include=dev
npm run build
npm run check

Then drive a workflow from the CLI:

# Validate one workflow file (direct paths do not need a .js extension).
node dist/cli.js validate ./path/to/workflow.js

# List personal and project workflows visible from a directory.
node dist/cli.js list ./path/to/project

# Run through the Codex SDK. Events are JSONL on stderr; the final result is
# JSON on stdout. The optional second argument is one JSON value exposed as `args`.
node dist/cli.js run ./path/to/workflow.js '{"files":["src/index.ts"]}'

# Resume a persisted Claude run (imported runs are read-only).
node dist/cli.js resume /path/to/claude/session/workflows/wf_id.json

# Serve over stdio, scoped to one project.
node dist/cli.js serve --stdio /path/to/project

# Serve over an authenticated loopback Streamable HTTP endpoint (URL + bearer
# token are printed once to stderr).
node dist/cli.js serve --http /path/to/project 0

Once served, both workflow_resume({ runId: "wf_..." }) and workflow_run({ resumeFromRunId: "wf_..." }) discover that Claude run inside the scoped project's Claude state. Use claudeRunPath only when duplicate historical metadata requires explicit selection.

Reading a complete result

Every newly completed service run stores one immutable UTF-8 result artifact. The compact workflow_run_status.run.result reference and the run.completed event both include its artifactId, media type, total UTF-8 byte count, line count, and SHA-256 checksum. When truncated is true, inline content is only a display prefix; it is not the complete result.

workflow_result_read accepts only a scoped runId plus that opaque artifactId—never a filesystem path. Pages default to 16 KiB and may request 4 through 65,536 bytes. Page ends are moved backward when necessary so concatenating content never splits a UTF-8 code point:

import { createHash } from 'node:crypto'

const statusCall = await client.callTool({
  name: 'workflow_run_status',
  arguments: { runId },
})
const status = statusCall.structuredContent as {
  run: {
    result: {
      artifactId: string
      checksum: { algorithm: 'sha256'; value: string }
    }
  }
}

const parts: string[] = []
let cursor: string | undefined
for (;;) {
  const call = await client.callTool({
    name: 'workflow_result_read',
    arguments: {
      runId,
      artifactId: status.run.result.artifactId,
      ...(cursor === undefined ? {} : { cursor }),
      maxBytes: 16_384,
    },
  })
  const { page } = call.structuredContent as {
    page: { content: string; hasMore: boolean; nextCursor?: string }
  }
  parts.push(page.content)
  if (!page.hasMore) break
  if (page.nextCursor === undefined) throw new Error('missing continuation cursor')
  cursor = page.nextCursor
}

const completeResult = parts.join('')
const digest = createHash('sha256').update(completeResult, 'utf8').digest('hex')
if (digest !== status.run.result.checksum.value) throw new Error('result integrity mismatch')

String results are raw text/plain; objects, arrays, numbers, booleans, and null are pretty printed application/json; JavaScript undefined is the text/plain bytes undefined. An empty string has zero bytes and zero lines. A top-level string containing a lone UTF-16 surrogate fails before completion because it has no lossless UTF-8 representation. Non-terminal runs return result-not-ready; failed, cancelled, or interrupted runs return result-unavailable; a completed legacy run without an artifact also returns result-unavailable; and missing retained bytes return result-expired. Malformed, stale, or non-UTF-8-boundary cursors return invalid-cursor.

FileWorkflowStore retains result bytes with the run directory and defaults to a 64 MiB result ceiling. Configure maxResultBytes when constructing the store if the host needs a different bounded policy (up to the hard 512 MiB safety ceiling). A result over that ceiling fails the run before run.completed rather than publishing another irreversible prefix. The direct workflow-mcp run CLI still writes its full result to stdout; the paginated contract applies to durable service/MCP runs.

Embedding

The public API uses plain names and hands the host full control of the MCP server, transport, and authentication lifecycle:

import {
  CodexAgentProvider,
  FileWorkflowStore,
  WorkflowService,
  registerWorkflowMcpTools,
} from 'workflow-mcp'

const service = new WorkflowService({
  store: new FileWorkflowStore('/private/application/state/workflows'),
  provider: () => new CodexAgentProvider({
    codexPathOverride: '/approved/codex',
    // Required before a host may attest that normal user/project MCP servers cannot leak into
    // an automatically replayed read-only workflow attempt.
    configurationIsolation: {
      codexHome: '/private/application/state/workflow-codex',
      authenticationFile: '/home/user/.codex/auth.json',
      // This must come from inspection of the exact executable plus user/project/system/managed
      // configuration layers. Omit it and use "unknown" below when the host cannot prove that.
      effectiveConfigurationFingerprint: verifiedCodexPolicyDigest,
    },
    capabilities: { inheritedMcpServers: 'disabled' },
  }),
  sandbox: { mode: 'read-only', approvalPolicy: 'never', network: false },
})
await service.initialize()

// The host still owns McpServer and its transport/authentication.
registerWorkflowMcpTools(mcpServer, service, { cwd: projectDirectory, clientId: sessionId })

Documentation

• docs/ARCHITECTURE.md — the full technical reference: the pinned Claude-workflow compatibility profile, the exact runtime realm, discovery/precedence, cache and resume mechanics, the MCP architecture, the Codex SDK findings, and the conformance matrix.
• docs/EXECUTION_PLAN.md — the phased build decisions.
• docs/RELIABILITY_IMPLEMENTATION_PLAN.md — the unattended-execution and failure-domain plan.

Agent Code and standalone use

workflow-mcp began as a feature of Agent Code — an open-source Electron IDE for driving the real Claude Code and Codex CLIs across a multi-agent workspace. Agent Code embeds this runtime through its existing MCP host and renders each run as a live, first-class feed card: phases and agents as vertical lists, with prompt, activity, and outcome expandable inline.

That embedded path remains supported, but it is no longer required. The Docker-first product above adds the supervised owner, project-scoped Codex MCP proxy, terminal and optional browser clients, credential and authoring controls, offline maintenance, and verified release machinery needed to run without the desktop application. Generic OCI/MCP-registry mode is intentionally session-bound; use the checksummed Compose bundle when runs must outlive an MCP client connection.

Status

The loader, execution runtime, durable service, MCP facade, Agent Code embedding, and Docker-first standalone implementation are in place. The first stable container release still requires the external platform qualification, protected release controls, and registry publication documented in the standalone implementation ledger. Compatibility is pinned to an observed Claude Code profile — a snapshot of a fast-moving upstream, not a promise about future versions.

License

MIT