
Gives Claude four operations to manage YAML workflow playbooks: list permanent workflows by category or tag, retrieve a specific workflow with global instructions prepended, create new permanent workflows in the library, and write temporary one-shot workflows. Built on the mcp-ts-core framework with a filesystem-backed in-memory index, semver-aware versioning, and automatic rebuilds when workflow files change. Includes tools enumeration so you can see which MCP servers a workflow depends on before running it. Useful when you're orchestrating multi-step agent tasks and want structured, reusable playbooks instead of freeform prompting. Runs over stdio or streamable HTTP with no API keys required.
Store, query, and create YAML workflow playbooks for LLM agents via MCP. STDIO or Streamable HTTP.
A declarative workflow library for LLM agents, backed by local YAML files. Store, list, and retrieve named, versioned multi-step playbooks — each a sequence of MCP server/tool calls — for permanent reuse or one-shot temporary runs. Runs as a stdio process or a local Streamable HTTP server; an in-memory index rebuilds automatically as files change.
| Tool | Description |
|---|---|
workflow_list | List all permanent workflows in the index, with optional keyword, category, and tag filters. |
workflow_get | Retrieve a complete workflow definition by name, with global instructions prepended. |
workflow_create | Write a new permanent workflow YAML to the library. |
workflow_create_temp | Write a temporary workflow draft, indexed but excluded from list results, kept until deleted. |
workflow_delete | Remove a permanent workflow or a temporary draft by name and optional version, after the user confirms the resolved target. |
workflow_list toolquery (name and description), category (substring), and tags (all must match) filters, case-insensitive; temporary drafts are never listednotice echoing the applied filtersincludeTools: true adds each workflow's unique server/tool pairsworkflow_get toolname plus an optional semver version; omit version for the highest available, drafts includedsource (permanent or temp), and globalInstructions from global_instructions.md (null when absent); fails with not_found or version_not_found{{input.foo}} come back verbatim — the server never interpolates themworkflow_create toolname, semver version, description, author, category, and steps (plus optional tags); stored as one file per name@version under categories/<category>/, with created and updated dates stampedkey (name@version) and filePath; fails with already_exists when a permanent workflow or a draft holds the key, invalid_input, or write_failedworkflow_create_temp toolworkflow_create fields minus category; stored under temp/, retrievable with workflow_get, never listed, and kept until workflow_delete removes itstatus is created, or overwritten when it replaced the draft of the same key (keeping that draft's created_date); a key held by a permanent workflow fails with already_existsworkflow_delete toolname plus an optional version (omitted: the highest across permanent workflows and drafts); the user confirms the resolved name@version, source, and file before anything is deleted, so a client without elicitation cannot deletesource, plus a notice when another file declaring the same key now takes its place; fails with cancelled, confirmation_invalid (a prompt never issued, already answered, older than 10 minutes, or issued to another caller), target_changed (the file or its content changed after the user saw it), or not_foundBuilt on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.
Workflow library:
categories/ and temp/ are validated at index time; an invalid file is skipped and logged, never crashes the servername@version, read from each file's content rather than its name and kept fresh by a debounced recursive watcher; a permanent workflow outranks a draft with the same keyv, surrounding whitespace, and build metadata are dropped, so v1.0.0+build.5 is stored, keyed, and looked up as 1.0.0name@version, so keys whose slugs coincide never share a file_index.json snapshot is written on every rebuild for external tooling and debuggingAgent-friendly output:
source: "permanent" | "temp" on every workflow_get response and typed reason codes (not_found, version_not_found, already_exists, cancelled, confirmation_invalid, target_changed, index_unavailable, …) on failures, so callers branch on data instead of parsing error stringsworkflow_get always returns globalInstructions alongside the workflow definition in the same responseworkflow_list's optional includeTools flag pre-derives the unique server/tool pairs used by a workflow, and an empty result echoes the applied filters with a broadening hint instead of returning nothingNo API keys required. The server reads from a local workflows-yaml/ directory by default.
Add the following to your MCP client configuration file:
{
"mcpServers": {
"workflows-mcp-server": {
"type": "stdio",
"command": "bunx",
"args": ["@cyanheads/workflows-mcp-server@latest"],
"env": {
"MCP_TRANSPORT_TYPE": "stdio",
"WORKFLOWS_DIR": "/absolute/path/to/your/workflows-yaml"
}
}
}
}
Or with npx (no Bun required):
{
"mcpServers": {
"workflows-mcp-server": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@cyanheads/workflows-mcp-server@latest"],
"env": {
"MCP_TRANSPORT_TYPE": "stdio",
"WORKFLOWS_DIR": "/absolute/path/to/your/workflows-yaml"
}
}
}
}
Or with Docker:
{
"mcpServers": {
"workflows-mcp-server": {
"type": "stdio",
"command": "docker",
"args": [
"run", "-i", "--rm",
"-e", "MCP_TRANSPORT_TYPE=stdio",
"-v", "/absolute/path/to/your/workflows-yaml:/workflows-yaml",
"-e", "WORKFLOWS_DIR=/workflows-yaml",
"ghcr.io/cyanheads/workflows-mcp-server:latest"
]
}
}
}
For Streamable HTTP, set the transport and start the server:
MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp
The repository ships a workflows-yaml/ directory with example workflows organized under categories/. These are ready to use as a starting point. The workflows-yaml/global_instructions.md file contains instructions the server prepends to every workflow_get response — edit it to set global guidance for your agent.
workflows-yaml/ seed).git clone https://github.com/cyanheads/workflows-mcp-server.git
cd workflows-mcp-server
bun install
cp .env.example .env
# edit .env if needed — most settings have defaults
| Variable | Description | Default |
|---|---|---|
WORKFLOWS_DIR | Absolute or relative path to the workflows root directory. | ./workflows-yaml |
GLOBAL_INSTRUCTIONS_PATH | Path to the global instructions markdown file. Derives from WORKFLOWS_DIR when not set. | <WORKFLOWS_DIR>/global_instructions.md |
WATCHER_DEBOUNCE_MS | Milliseconds to debounce filesystem change events before rebuilding the index. | 500 |
MCP_TRANSPORT_TYPE | Transport: stdio or http. | stdio |
MCP_HTTP_PORT | Port for HTTP server. | 3010 |
MCP_SESSION_MODE | HTTP sessions: auto or stateful. workflow_delete's confirmation prompt needs a live session, so HTTP startup with stateless fails with a configuration error. Ignored over stdio. | stateful, declared in src/index.ts |
MCP_AUTH_MODE | Auth mode: none, jwt, or oauth. | none |
MCP_REQUEST_STATE_KEY | Seals the requestState a confirmation round carries (≥ 32 bytes, the same on every instance, e.g. openssl rand -base64 32); any state the server did not seal is rejected before the handler runs. workflow_delete's round carries its confirmation id, so set it for HTTP deployments. | unset |
MCP_LOG_LEVEL | Log level (RFC 5424). | info |
LOG_TOOL_FAILURE_PAYLOADS | Log each failed tool call's arguments and result, redacted by key name and capped at LOG_TOOL_FAILURE_PAYLOAD_MAX_BYTES (default 16384). A secret inside a free-form value is not redacted. | false |
OTEL_ENABLED | Enable OpenTelemetry instrumentation (spans, metrics, completion logs). | false |
See .env.example for the full list of optional overrides.
Build and run:
# One-time build
bun run rebuild
# Run the built server
bun run start:stdio
# or
bun run start:http
Run checks and tests:
bun run devcheck # Lint, format, typecheck, security
bun run test # Vitest test suite
bun run lint:mcp # Validate MCP definitions against spec
docker build -t workflows-mcp-server .
docker run --rm \
-v /path/to/workflows-yaml:/workflows-yaml \
-e WORKFLOWS_DIR=/workflows-yaml \
-p 3010:3010 \
workflows-mcp-server
The Dockerfile defaults to HTTP transport, stateful session mode (required — see MCP_SESSION_MODE above), and logs to /var/log/workflows-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.
| Directory | Purpose |
|---|---|
src/index.ts | createApp() entry point — registers tools and inits the workflow index service. |
src/config/ | Server-specific environment variable parsing and validation with Zod. |
src/mcp-server/tools/ | Tool definitions (*.tool.ts). |
src/services/workflow-index/ | WorkflowIndexService — YAML parsing, index build, watcher, semver lookup, write helpers. |
tests/ | Unit and integration tests mirroring src/. |
workflows-yaml/ | Seed workflow library — categories/ for permanent workflows, temp/ for temporary drafts, global_instructions.md for agent-global guidance. |
See CLAUDE.md for development guidelines and architectural rules. The short version:
try/catch in tool logicctx.log for request-scoped loggingsrc/mcp-server/tools/definitions/index.tsWorkflowIndexService, not directly in tool handlersIssues are welcome. Run checks and tests before submitting:
bun run devcheck
bun run test
Apache-2.0 — see LICENSE for details.
WORKFLOWS_DIRdefault: ./workflows-yamlAbsolute or relative path to the workflows root directory.
GLOBAL_INSTRUCTIONS_PATHPath to the global instructions markdown file. Derives from WORKFLOWS_DIR when not set.
WATCHER_DEBOUNCE_MSdefault: 500Milliseconds to debounce filesystem change events before rebuilding the index.
MCP_LOG_LEVELdefault: infoSets the minimum log level for output (e.g., 'debug', 'info', 'warn').
MCP_HTTP_HOSTdefault: 127.0.0.1The hostname for the HTTP server.
MCP_HTTP_PORTdefault: 3010The port to run the HTTP server on.
MCP_HTTP_ENDPOINT_PATHdefault: /mcpThe endpoint path for the MCP server.
MCP_AUTH_MODEdefault: noneAuthentication mode to use: 'none', 'jwt', or 'oauth'.