🧭 OriginSpan MCP

A field named C_ID deserves more context than a confident guess.

Explore what an application does, what talks to what, where fields appear, and how SQL transforms them. OriginSpan brings source references, recorded relationships and visible gaps to the investigation.

Use it with Copilot or another MCP client, or generate reports through the optional local-model workflow. The MCP evidence server itself does not require a model subscription.

Status: pilot prerelease. Working discovery tools and measured feasibility; business accuracy and production capacity still require qualification.

Get started · See the evidence · Explore the docs

Follow the evidence

A declaration is an observed fact. A proposed business meaning is an interpretation. OriginSpan keeps those distinctions visible; a valid citation alone does not establish business correctness.

Related MCP server: spine

Under the hood

flowchart LR
    S[Approved code, schemas and metadata] --> E[Extraction and versioned evidence index]
    E --> M[MCP tools]
    M --> C[Copilot or another MCP client]
    M --> L[Local controller with an approved model]
    C --> R[Cited findings and review]
    L --> R

Parsing records source facts. Retrieval gathers relevant evidence and relationships. The selected AI interprets business context. Unknowns remain unknown until more evidence or a reviewer resolves them.

The index persists between questions. One user prompt can orchestrate several bounded tool calls; the whole repository does not need to enter one model prompt.

Pick your route

Python

Use Python 3.12 or newer, preferably in a virtual environment:

python -m pip install originspan-mcp==0.5.2rc2
originspan --help

Register and analyze an approved repository. Replace the example path and APPLICATION_ID with the ID returned by registration:

originspan --state .originspan --project demo --root /absolute/path/to/repo register MyApp --path /absolute/path/to/repo
originspan --state .originspan --project demo --root /absolute/path/to/repo analyze APPLICATION_ID
originspan --state .originspan --project demo --root /absolute/path/to/repo serve --transport stdio

On Windows, use a path such as C:\Projects\MyApp. The last command starts the MCP server. Configure your client to launch it with the same state, project and trusted root; use the installed executable's absolute path if it is outside the client's PATH.

Windows offline bundle

Download the Windows x64 / Python 3.13 ZIP from Releases. It contains the application, dependency wheels, checksums and installation scripts.

Follow the pilot quickstart. Python, model weights and llama.cpp binaries are separate prerequisites for the full local-model workflow. The MCP evidence tools can be used without a reasoning model.

Docker

docker pull ghcr.io/kunaldrafts25/originspan-mcp:0.5.2rc2
docker run --rm ghcr.io/kunaldrafts25/originspan-mcp:0.5.2rc2 --help

The Linux amd64 image runs as a nonroot user and defaults to MCP stdio. It includes PostgreSQL client support; model assets are separate. See the container guide for persistent state and client setup. ARM64 and GPU/model execution require separate validation.

Give it a useful first question

After registration and analysis, try this in your MCP client:

Use OriginSpan MCP as the application evidence source.
In project demo, locate MyApp and use its latest completed analysis.

Explain the application, draw its architecture in Mermaid, and find
demographic-related fields. Explain representative SQL transformations,
value lineage and join/filter dependencies where evidence exists.

Follow relevant evidence and result pages. Cite file locations.
Separate confirmed source facts from interpretations and unresolved meaning.
Report coverage gaps. Do not modify files or approve mappings.

A remote server needs an approved source snapshot of its own. Supplying a laptop path does not give that server access to your files.

Take the local route

The optional controller uses approved GGUF model files and a managed llama.cpp runtime. After provisioning, this path can retrieve evidence and generate reports without an online model provider, with no online fallback.

originspan --state .originspan --project demo discover APPLICATION_ID --analysis ANALYSIS_ID --profile laptop-qwen-v1

Create and verify the profile first using the local model guide. Profiles pin asset hashes, runtime versions and budgets. The reported POC used unchanged model weights; no fine-tuning was performed.

The numbers, with their context

Recorded local validation used the public Bank of Anthos sample and Qwen3.5-2B Q4_K_M:

Compared with our earlier local pipeline on the same pinned source and weights, reported input fell 70.9% and output fell 63.74%. Compact inputs remove repeated bookkeeping; known parser facts can be rendered directly.

• These are internal workflow savings, not a measured Copilot billing or answer-quality comparison.

• Hidden reasoning usage was unavailable. Earlier and final timing environments differed.

• Cold validation took 231.353 seconds, including indexing and startup; model download was excluded.

• Neural embeddings were disabled for this run.

• The bounded summary collected 12/60 ORM and 12/124 documentation records. Twelve meaning proposals are not complete business classification.

• Citation-reference validity passed; business correctness still requires review.

• Release packaging was separately checked with a fresh isolated installation, sample indexing and real MCP stdio. The public repository's local suite recorded 149 passed and one skipped external PostgreSQL integration test.

Read the report, raw proof and limitations.

Useful boundaries

Dynamic behavior, unsupported formats and missing context can leave gaps. A larger model or faster machine needs its own accuracy and capacity evaluation; the small banking proof does not establish terabyte performance or production concurrency.

This workflow uses code, schemas and approved metadata. It does not profile live customer records, execute application code, automatically edit sources or write to target systems. Interpretations cannot approve themselves.

Treat source text as untrusted evidence, never instructions. Keep credentials and customer records out of issue reports.

Trying OriginSpan?

Share your experience to let the maintainer know what you tried, whether it worked, and what would make it more useful. Installation feedback and small experiments are welcome; you do not need a production deployment. Sharing your organization or a public project link is optional. Reports are public, so leave out private code, credentials and customer data.

The badges above show monthly PyPI downloads and downloads of GitHub release assets. They count download events, including repeat downloads and automation; they do not identify people or establish active users. New-package statistics can be delayed, and badges are cached.

Maintainers can see recent repository visits and clones in Insights → Traffic. See project activity metrics for the meaning and limits of each signal.

Keep exploring

Good issue reports include the version, language/framework, sanitized source pattern, expected evidence and actual result. A tiny reproducible example beats a mystery screenshot.

Built by Kunal Singh

Created and maintained by Kunal Singh.

Copyright 2026 Kunal Singh. Licensed under Apache-2.0. See NOTICE and third-party notices. Dependencies, model/runtime assets and public samples retain their own licenses and copyright notices.