The file-native knowledge base

Ask what your sources already know.
Keep the evidence attached.

OmniTome allows people and teams to turn real source material into sealed, searchable Knowledge Bases, as well as serve them to any MCP client. The local path is one file, possession is the entitlement, and retrieval returns citations.

Install OmniTome See how it flows
  • Local-first and file-native
  • Any MCP client over stdio
  • Deterministic, cited retrieval
illustrative retrieval result deterministic retrieval
>

How do we prepare and deploy a service release?

decision sufficient
search exact
breadth 0.5

Validate the release candidate, publish the artifact, then run the deployment workflow for the target environment.

01

Repository / release workflow

current revision · deployment scope
sha256:7be2...
02

Wiki / release runbook

current revision · operations scope
sha256:31fd...
served from team.omnitome · policy 7 · breadth-v1

Why this exists

Useful answers rarely live in one source.

Operational knowledge spans repositories, wikis, work-item trackers, diagrams, file shares, and local documents. OmniTome records each observation's locator, revision, scope, and content hash before any knowledge is derived from it.

Compile seals exactly what one grant permits into a queryable SQLite file. Retrieval is plain code; a client model may turn the cited result into prose, but the evidence and its access boundary remain in the artifact.

Desktop, terminal, or assistant

Your knowledge.
Your way to work.

Install with the Python you already have. Collect, build, and ask in the desktop, automate with the CLI, or connect your preferred MCP client.

Desktop

A shortcut to your workspace

The Windows installer adds desktop and Start Menu shortcuts. Light and dark themes keep sources, build progress, and answers in one consistent workspace.

CLI + MCP

The same files, more ways in

A Python wheel installs the command-line tools and desktop together. Use scriptable stages or serve a sealed Knowledge Base to your assistant.

Evidence

Inspect what supports an answer

Optional model checks assess each proposed claim against its selected excerpts and check whether the draft covers the question. Exact source quotes remain available to inspect.

Read what the checks establish and where they stop

The five-verb path

Files connect every stage.

Scrape near the source, extract near the model, compile without either, and serve where the reader works. The separate flow page draws the people and files.

  1. 01

    Scrape

    Collect configured sources and record provenance.

    .omniblob
  2. 02

    Extract

    Propose typed knowledge; reject claims without support.

    model-assisted extraction
  3. 03

    Compile

    Seal exactly what one grant permits.

    .omnitome
  4. 04

    Embed

    Add a named semantic-search space.

    optional overlay
  5. 05

    Serve

    Open one or several files over MCP.

    omnitome-mcp
Follow documents and people through the pipeline

Optional .omniblobx and .omnitomex envelopes compress the byte-unchanged inner artifact. The current wrap command issues unsigned, unencrypted envelopes and reports that state explicitly.

Named feature

End-to-end incremental updating.

Returning operators pay for source changes, not another complete build.

01

Scrape the delta

Use revision- and date-aware enumeration to fetch only moved documents, and resume admitted revisions after interruption.

02

Extract the delta

Enrich only changed documents, with one durable journal record per document and mid-run resume.

03

Update the seal

compile --update swaps changed rows and computes retired atoms against the Knowledge Base.

04

Fill missing vectors

Embedding schedules only content hashes the stored space does not already hold.

Compatible checkpoints, extraction journals, and embedding identities allow unchanged work to be reused. Scrape progress shows source position, active container, exact fractions, bytes, and a proportional bar; bounded GET retries honor Retry-After, and stdout remains one JSON document.

Source families

Collect where the facts are written.

Credentials authenticate the reader and never travel in the archive. Every connector carries the observed scope and revision forward.

01

Azure DevOps

Repositories, project wikis, and work items.

azure-devops
02

GitHub

Explicit repositories, wikis, and issue threads.

github
03

Confluence

Selected spaces and pages.

confluence
04

SharePoint

Document libraries and modern site pages through Microsoft Graph.

sharepoint
05

Figma

One configured file, with frames, text, and supported diagram relationships.

figma
06

Jira Cloud

Explicit projects and complete visible issue threads. Hermetic and best-effort; not live verified.

jira
07

Local files

Documentation, code, diagrams, and office files on this machine.

filesystem

Software source coverage

From interface to deployment.

Preserve web and desktop UI source beside application code, RPC schemas, SQL and Entity Framework declarations, and infrastructure definitions. Readers retain authored evidence without running the application.

JSX & TSXVue & SvelteAstro & CSS XAML & Qt UIProtocol Buffers & GraphQLSQL & EF Bicep & ARMTerraform & KubernetesHelm

Built-in readers cover these source families. The optional language pack adds sixteen grammar families, including Java, Go, Rust, Swift, and QML. Unsupported structures remain source text or report partial analysis; static declarations do not establish runtime behavior.

Read the coverage matrix and parser limits

MCP over standard input and output

One file plugs into any MCP client.

The client launches omnitome-mcp local as a child process. There is no daemon and no authentication exchange, because compile already decided what the file contains.

omnitome mcp-config emits installed executable and absolute paths for this machine. The repository also carries a placeholder client block.

Open the sample MCP client configuration
mcp client configuration
{
  "mcpServers": {
    "omnitome": {
      "command": "/absolute/path/to/omnitome-mcp",
      "args": [
        "local",
        "--kb", "/absolute/path/to/kb.omnitome",
        "--space", "ollama-embed",
        "--inference-config", "/absolute/path/to/inference.json"
      ]
    }
  }
}

Install and quickstart

Build the file.
Ask the question.

The Windows kit adds a desktop shortcut using your installed Python 3.12+. A wheel also installs the CLI and UI on other platforms; the desktop needs Tk. Follow the fast path to configure a source and build your first Knowledge Base.

Follow the fast path
First Knowledge Base 
# Install a built distribution, then configure one source.
uv tool install './omnitome-<version>-py3-none-any.whl[production]'
omnitome configure producer.json

# Build and serve the sealed file through the five verbs.
omnitome scrape producer.json -o raw.omniblob
omnitome extract raw.omniblob -o enriched.omniblob \
  --file-policy file-policy.yaml --schemas config/extraction-schemas \
  --inference-config inference.json
omnitome compile enriched.omniblob -o kb.omnitome \
  --grant-profile everything --policy policy.json \
  --file-policy file-policy.yaml
omnitome embed kb.omnitome --space ollama-embed \
  --inference-config inference.json
omnitome-mcp local --kb kb.omnitome --space ollama-embed \
  --inference-config inference.json

# ... or to query the same file directly without MCP:
omnitome query kb.omnitome "How do we deploy this service?" \
  --space ollama-embed --inference-config inference.json

InstallBuild, install, upgrade, and remove the distribution. Security postureRead the fail-closed boundaries, threat model, and reporting path. DocumentationOpen the complete guide and runbook index.