Categories
Self-Hosted MCP Server Knowledge Graph Data Catalog Data Quality AI Data Governance Dataset Management Metrics LayerStars
Forks
Watchers
Developer links
Datakeel
Datakeel is a workbench where teams proofread AI training data, trace every business number back to its definition, and approve risky changes before anything is deleted. Evidence review marks each claim as present, absent, or unknown, so missing annotations never pass as clean, and accepted decisions are stored apart from the extracted evidence. Ask data answers a question like last quarter's retail revenue from registered metrics, shows the bound SQL, source tables, and time range, and refuses or asks for clarification when a question is ambiguous. A data catalog maps sources, tables, fields, owners, lineage, and glossary terms, and an asset market handles access requests with purpose and expiry. A knowledge graph review queue turns document extractions into entities and relations that stay candidates until a person checks the source sentence. Dataset versions freeze file manifests with SHA-256 hashes, quality scores rank samples, and workflows run health checks, cleanups, and metric changes that pause at an approval sheet bound to a digest of the plan. Built-in agents and an MCP server can propose actions but never apply them alone, every receipt is kept in an append-only history, and the interface ships in English and Chinese with three themes. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. Apache-2.0 licensed.
Benefits
- Honest Evidence, Never Guessed
- Each claim is present, absent, or unknown, so missing annotations stay visibly unresolved instead of silently counting as negatives when you review samples and build dataset versions.
- Numbers You Can Trace
- Metric answers cite their definition version, owner, parameterized SQL, source columns, and time range, and ambiguous questions get a clarification or a refusal instead of a guess.
- Humans Approve Every Change
- Deletions, metric edits, and outbound transfers become frozen plans with expiry and impact notes, and nothing runs until a person signs the digest-bound approval sheet.
- Starts Small, Scales Up
- The offline demo needs no database server or model key, and the same app moves to PostgreSQL and S3 behind an authenticating reverse proxy for a small team.
Features
- Evidence Proofreading Workspace
- Review images, captions, and annotation matrices with highlighted anchors, filter for gaps like unannotated pedestrians, accept assertions, and create versions from selected samples.
- Ask Data Answers
- Ask business questions in plain language and get a metric value with its definition, SQL, and time interval, plus related document claims and a replayable saved answer.
- Catalog And Lineage
- Browse sources, tables, fields, and owners, check which metrics use a table, inspect upstream and downstream impact, and import metadata from outside systems.
- Knowledge Graph Review
- Inspect extracted relations with their evidence sentence and verification status, accept or reject them in bulk, and view the entity graph, extraction runs, and schema proposals.
- Workflows And Approvals
- Run dataset health checks, low-quality cleanup, and metric changes step by step, then review pending approvals with objects, impact, and an irreversible-action notice.
- Agent And MCP Access
- Chat with a controlled-action agent or connect an MCP server that proposes plans, tracks conversation budgets, and records traces and costs for every call.