reducto-beta is Reducto’s private-beta MCP server. It gives an authenticated coding agent access to Studio pipelines and Contextual Schema Optimization workflows.
The name reducto-beta is the label you configure in your MCP client. The hosted endpoint is:
https://studio.reducto.ai/api/mcpConnect
claude mcp add --transport http reducto-beta --scope user https://studio.reducto.ai/api/mcpRun /mcp inside Claude Code, select reducto-beta, and complete the browser sign-in.
codex mcp add reducto-beta \
--url https://studio.reducto.ai/api/mcp \
--oauth-resource https://studio.reducto.ai/api/mcp
codex mcp login reducto-beta --scopes mcpAdd a remote MCP server named reducto-beta:
{
"mcpServers": {
"reducto-beta": {
"url": "https://studio.reducto.ai/api/mcp"
}
}
}Your client should open the Reducto authorization flow when it first connects.
Authentication uses your Reducto Studio account and organization. The server applies the same pipeline access checks as Studio. Switching organizations requires reconnecting or reauthorizing the MCP session.
Recommended workflow
For Contextual Schema Optimization tasks, the coding agent should:
- Inspect the current schema or extraction goal and skim representative documents.
- Explain its understanding of the documents, desired output, and downstream use.
- Prepare
setup_use_caseoroptimize_schemawithconfirm_and_submit: false. - Show the final inputs and warnings, then ask for explicit confirmation.
- Submit with
confirm_and_submit: trueonly after confirmation. - Repeatedly call
get_optimizer_runwith the response’spoll_inputuntil the run reaches a final status. - Read the report from the response and surface every unresolved question, ambiguity, and assumption before recommending the configuration.
Ground truth is optional. If it is missing, the agent should ask once whether expected outputs exist and where to find them, but it should not block the run solely because ground truth is unavailable.
Tool catalog
Context and pipeline discovery
| Tool | Purpose | Important inputs |
|---|---|---|
get_studio_context |
Show the authenticated organization, Studio deployment, and available tools | None |
list_pipelines |
Search, sort, filter, and paginate accessible pipelines | search, page, pageSize, sortField, sortDirection, filterProcessorTypes, excludeDemo |
get_pipeline |
Return a pipeline with ordered processors and working configurations | pipelineId |
Pipeline management
| Tool | Purpose | Important inputs |
|---|---|---|
create_pipeline |
Create a pipeline from PARSE, SPLIT, EXTRACT, EDIT, or CLASSIFY processor specifications |
name, processors[] |
update_pipeline |
Rename a pipeline or update processor names/configurations | pipelineId, optional name, processors[] |
upload_file |
Import one HTTPS file or prepare a local-file upload | pipelineId, name, sourceUrl or localPath, optional contentType |
upload_files |
Import or prepare uploads for up to 25 files independently | pipelineId, files[] |
run_pipeline |
Run all terminal processors for existing Studio file IDs | pipelineId, fileIds, optional completion wait controls |
create_share_link |
Create a public link to the current pipeline result snapshot | pipelineId, optional expiry: 1h, 1d, 1w, or never |
The server intentionally does not expose pipeline or processor deletion. update_pipeline leaves omitted processors unchanged.
Use these tools only when you intend to operate on a Studio pipeline. Do not create a temporary pipeline for Contextual Schema Optimization setup or schema optimization.
Contextual Schema Optimization
| Tool | Purpose | Important inputs |
|---|---|---|
setup_use_case |
Create a schema and prompt from at least three representative documents and an extraction goal | documents or workspace_ref, requirements, confirm_and_submit |
optimize_schema |
Improve an existing schema/prompt from at least three representative documents | current schema/prompt, documents, optional ground truth, confirm_and_submit |
health_check |
Coming soon; not available during the current design-partner beta | Do not invoke yet |
get_optimizer_run |
Poll one accepted setup or optimization run | run_id, optional workflow |
create_optimizer_workspace_upload |
Create a presigned upload for a native Contextual Schema Optimization workspace ZIP | workflow, document counts, result controls |
The health_check tool contract may change before release. Reducto will update this page when it is ready to use.
Provide documents
The MCP accepts three input patterns.
Local files
Use local_documents when the coding agent can read the files directly:
{
"local_documents": [
{
"id": "resume-1",
"name": "resume-1.pdf",
"local_path": "/absolute/path/resume-1.pdf",
"size_bytes": 48213
}
]
}For optimization, pair expected JSON through local_ground_truth and set document_id to the matching document. A local MCP client packages and uploads the workspace automatically. A hosted client returns a ZIP manifest and presigned upload instructions for the coding agent to complete.
HTTPS references
Use document_refs for public or presigned HTTPS URLs. Setup uses url; optimization uses document_url and optional ground_truth_url. Do not mix HTTPS references with local files in one preparation call.
Native workspace ZIP
Use create_optimizer_workspace_upload when the corpus exceeds 150 MiB, contains many files, or requires transforming a table of expected outputs into per-document ground_truth.json files. Upload the returned workspace, then pass its workspace_ref to setup_use_case or optimize_schema.
Example prompts
Create a resume extraction configuration
Use reducto-beta to set up extraction for the resumes in
/path/to/resumes. I need candidate contact details, work history, education, and skills. Preserve responsibility text verbatim. Inspect a few files, tell me your understanding, and prepare the run for confirmation.
Improve an existing configuration
Use reducto-beta to optimize this resume extraction schema against
/path/to/evaluation-set. Expected JSON is beside each PDF. Focus on missing employment dates and inconsistent responsibility arrays. Do not submit until I confirm the prepared run.
Results and safety
get_optimizer_run returns accepted, running, completed, partial, or failed. The agent must continue polling through intermediate states and should not present an answer until the run reaches a final status or four hours have passed.
Completed and partial runs include either safe inline artifacts or a temporary result bundle. The agent should inspect the report when instructed and distinguish measured ground-truth accuracy from estimated or consistency-based evidence.
reducto-beta does not expose pipeline deletion. Contextual Schema Optimization preparation is non-destructive, but submission creates a billable asynchronous run. The confirmation step exists to make that boundary explicit.
Troubleshooting
| Problem | What to do |
|---|---|
| The server is missing | Confirm the name is reducto-beta and the URL is https://studio.reducto.ai/api/mcp |
| Authentication expired | Reconnect the server or rerun your client’s MCP login flow |
| The wrong organization is active | Reauthorize using an account with access to the intended Studio organization |
The agent stopped at accepted or running |
Ask it to keep calling get_optimizer_run with the response’s poll_input |
| The MCP client cannot read local files | Use absolute paths and confirm the MCP client has filesystem access |
| The corpus is larger than 150 MiB | Use create_optimizer_workspace_upload and a native workspace ZIP |