Orchex

wundam/orchex

You'd reach for this when you need AI task orchestration that handles the complexity for you. It automatically breaks down requests into subtasks, runs them in parallel where possible, and routes each piece to the most appropriate model across six different LLMs. The self-healing aspect means it can recover from failures and retry with different strategies. Think of it as a smart middleware layer that sits between your application and various language models, deciding how to decompose work, which model handles what, and how to stitch results back together. Useful when you're building agents or workflows that benefit from multi-model strategies without manually coding all the routing and error handling logic.

tools: 12
auth: Secret needed
transport: STDIO
env vars: 5
last updated: Mar 13, 2026

View on GitHub

Tools

12 tools

Public tool metadata for what this MCP can expose to an agent.

initInitialize a new orchestration for a feature. Creates .orchex/active/manifest.yaml with streams.3 params

Initialize a new orchestration for a feature. Creates .orchex/active/manifest.yaml with streams.

Parameters* required

featurestring

Feature name (e.g., "user-authentication")

project_dirstring

Project directory (defaults to cwd)

streamsobject

Map of stream ID → stream definition

add_streamAdd a new stream to the active orchestration.9 params

Add a new stream to the active orchestration.

Parameters* required

depsarray

Stream IDs this depends on

idstring

Unique stream ID

namestring

Human-readable stream name

ownsarray

File paths this stream owns

planstring

Implementation plan for the stream

project_dirstring

Project directory

readsarray

File paths this stream needs to read (but does not own)

setuparray

Setup commands to run before the stream executes

verifyarray

Verification commands to run after the stream completes

statusGet the current orchestration status and progress.1 params

Get the current orchestration status and progress.

Parameters* required

project_dirstring

Project directory

executeRun the orchestration. Each call executes one wave of streams in parallel via LLM API, applies artifacts to the codebase, and runs verification commands. Call repeatedly for each wave, or use auto mode.7 params

Run the orchestration. Each call executes one wave of streams in parallel via LLM API, applies artifacts to the codebase, and runs verification commands. Call repeatedly for each wave, or use auto mode.

Parameters* required

dry_runboolean

Preview without executing (default: false)default: false

iterateboolean

When true with mode=auto, re-plan and re-execute if failures remain. Max 3 iterations.default: false

max_iterationsnumber

Maximum iterations for iterative mode (default: 3, max: 5).default: 3

modestring

auto: execute all waves sequentially. wave: execute one wave and return.one of auto · wavedefault: wave

modelstring

LLM model override (auto-detected based on provider)

project_dirstring

Project directory

wavenumber

Run a specific wave number (optional)

completeMark a stream as complete, or archive the entire orchestration.3 params

Mark a stream as complete, or archive the entire orchestration.

Parameters* required

archiveboolean

Archive the orchestrationdefault: false

project_dirstring

Project directory

stream_idstring

Stream ID to mark complete

recoverRecover streams stuck in in_progress state or failed state. Auto-detects stuck streams and resets them to pending for retry, or skips them.4 params

Recover streams stuck in in_progress state or failed state. Auto-detects stuck streams and resets them to pending for retry, or skips them.

Parameters* required

include_failedboolean

Also recover failed streams (not just stuck in_progress ones).default: false

modestring

retry: reset stuck streams to pending. skip: mark them as skipped.one of retry · skipdefault: retry

project_dirstring

Project directory

stream_idstring

Specific stream ID to recover. If omitted, recovers all stuck streams.

learnParse a planning document and generate stream definitions. Returns proposed streams for review before initialization.6 params

Parse a planning document and generate stream definitions. Returns proposed streams for review before initialization.

Parameters* required

deliverable_levelnumber

Header level for deliverables (2=H2, 3=H3)default: 2

document_contentstring

Plan document contents as string. When provided, used instead of reading document_path from disk. Useful for cloud mode where the server cannot access local files.

document_pathstring

Path to the markdown planning document

prefixstring

Optional prefix for stream IDs

project_dirstring

Project directory (for resolving relative paths)

validate_antipatternsboolean

Validate generated streams against anti-patternsdefault: true

init-planGenerate an annotated markdown plan template that teaches users how to structure documents for orchex learn. Creates a starter plan file with inline comments explaining what the learn pipeline parses.3 params

Generate an annotated markdown plan template that teaches users how to structure documents for orchex learn. Creates a starter plan file with inline comments explaining what the learn pipeline parses.

Parameters* required

feature_namestring

Feature name for the plan (used in title and example stream IDs)default: my-feature

output_pathstring

Output file path. Defaults to docs/plans/YYYY-MM-DD-<feature>-plan.md

project_dirstring

Project directory (defaults to cwd)

autoGenerate a plan from user intent, preview streams, and optionally execute. Flow: gather context → LLM generates plan → learn pipeline → preview → approve → execute → report.6 params

Generate a plan from user intent, preview streams, and optionally execute. Flow: gather context → LLM generates plan → learn pipeline → preview → approve → execute → report.

Parameters* required

approveboolean

If true, execute immediately after preview. If false (default), return preview only.default: false

approved_streamsarray

Stream IDs to approve for execution. Non-listed pending streams will be skipped. Omit this parameter to approve all streams (default behavior).

intentstring

What you want to build (e.g., "Add user authentication with JWT")

modelstring

LLM model override

project_dirstring

Project directory (defaults to cwd)

providerstring

LLM provider override

reset-learningReset learning data: thresholds, events, patterns, and/or reports.4 params

Reset learning data: thresholds, events, patterns, and/or reports.

Parameters* required

confirmboolean

Must be true to proceed

patterns_onlyboolean

Only reset patternsdefault: false

project_dirstring

Project directory

reports_onlyboolean

Only reset reportsdefault: false

rollback-streamRevert file changes made by a specific stream using git.2 params

Revert file changes made by a specific stream using git.

Parameters* required

project_dirstring

Project directory

stream_idstring

Stream ID to rollback

reloadRestart the MCP server to pick up rebuilt code. MCP client will reconnect automatically.

Restart the MCP server to pick up rebuilt code. MCP client will reconnect automatically.

No parameter schema in public metadata yet.

Configuration

ANTHROPIC_API_KEYsecret

Anthropic API key — auto-detected from shell env; override only

OPENAI_API_KEYsecret

OpenAI API key — auto-detected from shell env; override only

GEMINI_API_KEYsecret

Gemini API key — auto-detected from shell env; override only

DEEPSEEK_API_KEYsecret

DeepSeek API key — auto-detected from shell env; override only

ORCHEX_API_URL

Override cloud API URL (optional, defaults to https://orchex.dev)

Readme

Details

Repository: github.com/wundam/orchex
Package: @wundam/orchex
Transport: STDIO
Auth: Required
Registry: MCP registry, active