Datacron

Local MCP server to query and maintain a Markdown vault from Claude, Codex, Gemini, or another stdio MCP client, without sending the whole vault into the context.

Français | English

Datacron indexes a folder of Markdown notes, exposes a local MCP server, then returns the relevant notes or chunks to the client instead of a full dump. The vault stays an ordinary Markdown folder: Datacron only adds a .datacron/ sidecar for the index, logs, internal ULIDs, history, and the operation journal.

What is in place

Local measurement of the tool/impl pipeline actually received by the agent, 19 questions, 8k-token / 20-result configuration, July 17, 2026:

recall@5       0.89
recall@10      0.95
recall@20      0.95
MRR            0.73
nDCG@10        0.79
latency p50    57 ms
latency p95    276 ms
payload tokens 90567

The tool now matches the raw store at recall@5 (0.89): the previous gap came from globally comparing scores produced by separate AND and OR queries, not from the response budget or a BM25 limitation. Repair-on-read throttling brings its own p50 down to 0.009 ms; the first full sweep of the session remains visible in p95. The golden set does not yet contain a forbidden_paths case and the vault has no indexed supersedes relationship.

Installation

Windows: one double-click installer

The easiest way on Windows: download Datacron-Setup.exe from the latest Release, double-click it, and pick your vault. No Python, no terminal, no administrator rights; Datacron registers itself with your AI clients automatically. Full guide: Windows installation.

From source

From a clone of the repository:

python -m pip install -e ".[dev]"

Or, to install only the application:

python -m pip install -e .

Runtime prerequisites:

• Python 3.11+
• ripgrep available on the PATH for search_regex
• a folder of Markdown notes
• a supported stdio MCP client, such as Claude Desktop, Codex CLI, or Gemini CLI

Quick start

The easy path - one command detects your AI clients, initializes the vault, indexes it, and registers Datacron everywhere:

datacron setup            # interactive; add --yes for all defaults

See the installation guide for options (--client, --scope, writing, durability). Or step by step:

datacron init /path/to/vault
datacron index --vault /path/to/vault
datacron status --vault /path/to/vault
datacron mcp install --client claude-desktop --vault /path/to/vault

The mcp install subcommand above is dedicated to Claude Desktop. For Codex CLI, Gemini CLI, Antigravity, LM Studio, Cursor, and the other clients, use multi-client setup with datacron setup --client <identifier> or auto-detection with --client all.

Add to LM Studio

LM Studio 0.3.17+ has one user configuration and no project scope. The preferred command is:

datacron setup --yes --vault "VAULT_PATH" --client lmstudio --scope user

For a Python installation where datacron-mcp is on PATH, the equivalent read-only configuration can also be imported with this official deeplink:

Add to LM Studio

The link imports this example. Open LM Studio's MCP editor and replace both <YOUR_VAULT> placeholders before starting the server:

{
  "mcpServers": {
    "datacron": {
      "command": "datacron-mcp",
      "args": [],
      "env": {
        "DATACRON_VAULT_ROOT": "<YOUR_VAULT>",
        "DATACRON_READ_PATHS": "<YOUR_VAULT>",
        "DATACRON_DURABILITY": "best-effort"
      }
    }
  }
}

The example does not enable write tools. CLI setup is safer for packaged installations because it writes the actual executable path automatically.

Restart the configured client or clients after installation.

To run the server manually:

datacron mcp serve --vault /path/to/vault

The direct script entry used by the installer is also available:

datacron-mcp

datacron-mcp reads the vault from DATACRON_VAULT_ROOT.

Configuration

datacron init creates .datacron/VAULT.yaml. That file can carry vault-local configuration, notably query expansion:

query_expansion:
  supervision: [monitoring]
  sauvegarde: [backup]
  restauration: [restore]
  chiffrement: [encryption]
  sécurité: [security]
  validité: [validity]
  certificat: [certificate]

Useful environment variables:

Path lists use the OS separator (: on Unix, ; on Windows).

Writing

Writes are deliberately OFF by default. Without DATACRON_WRITE_PATHS, write tools return a clear error and create no file.

To enable writing to a specific subfolder:

$env:DATACRON_VAULT_ROOT = "G:\_DATA"
$env:DATACRON_READ_PATHS = "G:\_DATA"
$env:DATACRON_WRITE_PATHS = "G:\_DATA\_memory"
datacron mcp serve --vault G:\_DATA

datacron setup can also apply the allowlist machine-wide (user environment variable, opt-in) so every MCP client inherits it; default: _memory, _drafts, _journal. See the setup guide.

Available write tools:

• create_note_ai: creates a typed Markdown note, without overwrite.
• append_journal: adds an entry under a heading of an existing note.
• set_frontmatter: updates lifecycle fields and the rejected options list without modifying the Markdown body.
• patch_note_preamble: replaces or removes the Markdown preamble before the first ATX heading, with mandatory CAS control.
• patch_note_section: replaces the content under an existing heading with CAS control.
• delete_note_section: explicitly deletes an ATX H2-H6 section and its subtree.
• rename_note_section: renames only the title of an ATX H2-H6 section.
• revert_note: restores the exact bytes of a version kept in history.
• apply_organization_manifest: validates and then applies a local content-addressed bundle after confirmation bound to the exact admitted organization pre-state.

Guarantees:

• strict note confinement within DATACRON_WRITE_PATHS; organization-batch note sources and targets must also stay inside the unchanged live organization.scope and pass the live note-admission policy, including exclusions
• two internal exact-CAS targets for an organization batch: .datacron/VAULT.yaml, only to change the top-level organization mapping without changing organization.scope, and .datacron/ulids.json, only when Datacron derives the key migration required by a move_replace_exact
• atomic overwrite via temporary file + os.replace
• content-addressed history before modifying an existing note
• synchronous reconcile() after a normal write; immediate searchability is guaranteed only when reconciliation succeeds
• local audit log
• for an organization manifest: crash-consistent recovery and atomic replacement of each file; simultaneous visibility across several paths is not guaranteed

Concurrent multi-machine mode is not supported for writes: keep a single-writer rule on the vault.

For apply_organization_manifest, also stop every other Datacron client and server during the maintenance window. Before applying, keep a verified byte-exact backup outside the vault of the affected notes and the complete .datacron directory until every post-commit check is green. Call mode="validate" first, review the bounded hashes it returns, then reuse the exact confirmation_token with mode="apply". The token binds the manifest and payloads, all admitted Markdown notes inside organization.scope, the exact vault configuration and identity sidecars, and the projected report. It deliberately does not bind unrelated note bytes outside organization.scope. A change to any authenticated component invalidates the confirmation before mutation. history_mode=full is required at validation time. If Datacron derives identity-sidecar case-collision cleanup, also review identity_sidecar_case_canonicalization_count and its content-free SHA-256 before applying; both proofs are token-bound and retained in the durable receipt. An existing replace_exact or move_replace_exact source must carry its id in frontmatter; an identity available only from the sidecar is unsupported by this v1 schema. If the batch is already durably committed but index reconciliation or the planner oracle fails, the response says so explicitly (committed_index_incomplete or committed_report_mismatch) and the same call can be retried with the same token. An organization-batch blocker is reported by datacron ops inspect with a pending_batch_ reason and both single-note repair actions unavailable; use the full offline rollback procedure in the operational-health guide rather than repairing or quarantining one member.

MCP Tools

Reading

Writing

Operational

Advisory (experimental)

MCP resources:

• datacron://vault/map
• datacron://vault/info
• datacron://policy/active

Search

search_text combines several signals:

• FTS5/BM25 for the base lexical score
• FR↔EN query expansion configured in VAULT.yaml
• conservative temporal re-rank:
  • a note referenced in another note's supersedes is strongly demoted
  • confidence: low and confidence: needs_verification apply a light penalty
  • include_superseded=true brings historical notes back up

search_regex stays literal: it applies neither query expansion nor temporal re-rank.

Privacy and security

• Datacron does no telemetry.
• Datacron calls no cloud LLM.
• The MCP client, for example Claude, Codex, or Gemini, may send the chunks that Datacron returns to its provider. Datacron does not send it the full vault.
• Content returned to clients is wrapped in <vault_content>...</vault_content>.
• Results are bounded by count and by token budget.
• Filesystem access is confined by DATACRON_READ_PATHS and DATACRON_WRITE_PATHS.
• MCP operations are audited in the local logs.

CLI commands

datacron setup                      # guided path: init + index + client config
datacron setup --yes                # all defaults, no prompts
datacron setup --client all --scope both --vault /path/to/vault
datacron setup --protocol           # also install client memory rules
datacron protocol install --client all
datacron init /path/to/vault
datacron status --vault /path/to/vault
datacron index --vault /path/to/vault
datacron reindex --vault /path/to/vault
datacron scrub-init --vault /path/to/vault
datacron scrub --vault /path/to/vault
datacron eval --questions examples/eval-questions.example.yaml --vault /path/to/vault
datacron eval --questions local/golden.yaml --vault /path/to/vault --save-baseline
datacron eval --questions local/golden.yaml --vault /path/to/vault --compare --json
datacron mcp serve --vault /path/to/vault
datacron mcp install --client claude-desktop --vault /path/to/vault  # Claude Desktop only
datacron unregister --client all --scope both --vault /path/to/vault
datacron protocol uninstall --client all

Current limitations

• No vector search / embeddings: the spike is ruled out on the current golden because tool-level recall@5 at 0.89 matches the BM25 store. Re-evaluate if an expanded golden falls below 0.85 with the same evaluation.
• No autonomous agent: the MCP client orchestrates.
• No GUI.
• No concurrent multi-machine writes.
• Client detection in datacron setup is best-effort (a config directory or a binary on the PATH); an install in a non-standard location may be missed and can then be configured by hand.

Documentation

Full index: docs/en/index.md | Index français.

To get started:

• Installation and configuration guide
• Use Datacron with Ollama
• Frequently asked questions
• User guide

Technical references:

• Vault conventions (SPEC)
• Architecture and public surface
• Security boundary
• Integrity scrubber
• Operational health and durability
• Freshness contract

Development

python -m pip install -e ".[dev]"
ruff check .
ruff format --check .
mypy
pytest

License

Copyright 2026 Julien Bombled.

Licensed under the Apache License, Version 2.0.