Provenance is data
Spans survive extraction and retrieval, enabling verifiable line, page, region, and time citations.
The project behind the product
A deployment primitive for the agent web.
dir2mcp began as a Mistral Hackathon project and grew into a multi-provider, implementation-backed experiment in making private knowledge safely useful to agents.
The thesis
Private knowledge should be as easy to deploy as a web server.
Most valuable context already lives in repositories, archives, documents, and recordings. Uploading it all to a SaaS—or giving an agent raw filesystem access—should not be the only choices.
InputDirectory corpus
NormalizeExtract + chunk
GroundIndex + provenance
ServeMCP tools
Key decisions
Small surface. Deep internals.
The complexity stays behind stable tools so clients can search, inspect, and reason over many source types without learning the ingestion pipeline.
Capabilities compose
Embedding, OCR, speech, answers, and reranking can resolve to different local or cloud providers.
Coverage is honest
A skipped document or partial transcript stays visible instead of looking like a complete corpus.
Stable tool surface
Eleven tools cover the corpus lifecycle.
Core retrieval stays compact while optional media and generation tools reuse the same citation and error contracts.
| Tool group | Tools | Purpose |
|---|---|---|
| Retrieve | search, related, ask | Find passages, follow context, and generate cited answers. |
| Inspect | open_file, open_media_clip | Read the exact source slice behind a result. |
| Navigate | list_files, stats | Understand corpus contents, coverage, and model routing. |
| Transform | transcribe, annotate | Produce time-coded text and structured document data. |
| Compose | transcribe_and_ask, ask_audio | Chain media input and spoken output through the same evidence path. |
Open benchmark
Quality is a measured tradeoff.
The benchmark publishes its corpus construction, questions, raw responses, deterministic scorer, latency, and failure modes—not just one headline score.
contains-gold answer rate with the verifier off.
inline citation precision at file level.
unanswerable-question abstention when faithfulness verification is enabled.
The verifier improves refusal behavior but reduces correct answers and increases latency. Both runs and their limitations are published inthe benchmark report.
Built at a hackathon. Maintained like infrastructure.
The Dirstral team turned a Mistral-first prototype into a tested Go service, a canonical public specification, reproducible benchmarks, and a growing open-source contributor project.