claimbase
Summary: A read-only MCP server for querying a compiled, bitemporal claim graph.
recall: Answer a full question from the claim graph, ranked by similarity, provenance, currency, and supersession; returns claims with source, kind, assertion date, and interpreted intent.recalloptions: Useas_offor historical belief state,intentto override interpretation (mechanism | current | historical | neutral),include_supersededto include stale claims, andkto set result count.timeline: Show how belief about a subject changed over time, in assertion order, including which claims superseded which.conflicts: List open contradictions the compiler could not settle.stats: Report corpus counts by source, kind, and trust.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@claimbase is auto-promote still the promotion path?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
claimbase
Notes are a build artifact, not a source of truth.
A compiled knowledge substrate. The source of truth is an append-only capture log; a background compiler extracts atomic claims, tracks supersession, and emits a bitemporal claim graph that agents query over MCP. Nobody gardens.
The design is in DESIGN.md. What follows is what is actually built, and what the measurements say — including where they say it does not work.
Status: Phase 0, and honest about it
Compiled over the guru, guru-web and rellm repos — 3,287 events, 11,667 claims, 14,360 edges, from six adapters.
Against two baselines on nine gold questions:
ripgrep | chunk-RAG | claims | |
nDCG@10 | 0.229 | 0.500 | 0.679 |
mislead-rate | 0.500 | 0.333 | 0.000 |
mislead-rate is the fraction of questions where a superseded record outranks
every correct one. It is the number this project exists to move; nDCG mostly measures
whether the thing was findable at all.
These rest on nine questions and ranking weights adjusted against them. That is fitting to noise as much as to signal. The two claims that survive scrutiny are categorical rather than marginal: the mislead-rate does not depend on weight magnitudes, and some answers are unreachable by search at any quality — see below.
An answer-level bench (same model, same turn budget, recall versus ripgrep + file
reads, blind judging) ran three times: claimbase 3/4/3 correct, ripgrep 1/1/1. The
gap is robust; differences between successive ranking versions are not measurable at
that sample size, which is why tuning stopped.
Related MCP server: xrad
What it does that search cannot
Knows what stopped being true. A practice ends, eight documents go on describing it, and similarity gives the crowd the win. Ranking accounts for provenance, currency and whether a claim superseded others, so one correct claim can outrank eight stale ones — it wins with a lower cosine than the documents it beats.
Answers what was never written down. The sharpest failure in a real corpus is
not contradiction, it is silence: a practice abandoned with no record of the
abandonment. No compiler recovers that. assert captures it, and the answer becomes
reachable — in testing, ripgrep replied "the evidence does not contain any
information regarding..." to a question claimbase answered correctly.
Reads tables. A benchmark report is numbers, which carry almost no semantic signal; both baselines score 0.000 on metric questions. Compiled to claims, they become retrievable.
Design notes worth the detour
The adapter seam — six adapters were written and
core/was never touched. An abstraction with one implementation is not one.Trust tiers — adapters declare who authored a claim; core enforces what that permits. Agent-written prose cannot become a
fact. Claims that outrun their trust are demoted rather than dropped, keeping the evidence and withholding the authority.capability vs practice — a tool that still works but is no longer used has not been falsified. Conflating the two cost the same test score three times through three different rules before the label at the source was found to be wrong.
findings.md — the running log, including every measurement that came out badly and several instruments that turned out to measure nothing.
Running it
docker compose up -d # postgres + pgvector on 5433
docker compose exec -T db psql -U claimbase -d claimbase -f /migrations/001_init.sql
python -m claimbase import # compile a corpus
python -m claimbase embed # vectors (needs ollama)
python -m claimbase supersede && python -m claimbase resolve
python -m claimbase recall "is auto-promote still the promotion path?"Every command that can destroy work takes a verified snapshot first
(scripts/backup_db.sh, ~/claimbase-backups/), and refuses to run if the snapshot
fails. Extraction costs hours of GPU and was lost twice before that existed — once
to an unscoped delete, once to a TRUNCATE CASCADE that reached claims through a
foreign key. "Derived data is rebuildable" is not the same as "free to lose".
Extraction needs a local model server; everything else needs Postgres and Ollama.
Corpora are declared in corpora/*.toml — named, and never derived from
the working directory.
The memory layer — session ingest, memory PRs, usage measures — is specified in
GOALS.md and tested against a scratch database
(scripts/test_db.sh up), never the live one.
For agent access see MCP.md. Read-only by default; writes are opt-in, because a false claim written to the graph misleads every later session.
Licence
MIT
Available Tools
4 toolsconflictsB
Open contradictions the compiler could not settle. Empty is the normal state.
Inputs: limit.
recallA
Answer a question from the claim graph.
PASS THE QUESTION, NOT KEYWORDS. Phrasing carries meaning the ranking uses: "why is X still pending" is a question about mechanism and wants a durable explanation, while "is X still the current path" is about currency and wants the newest practice. Reduced to "X pending" both look identical and the wrong claims win. Full questions retrieve at least as well here — this is not a keyword index.
Returns claims ranked by epistemic standing — similarity, provenance, currency, and whether a claim superseded others — each with its source, kind, assertion date, and the intent the question was read as.
as_of: ISO date. Answers as the graph stood then ("what did I believe in June?"). intent: override if the reading looks wrong — mechanism | current | historical | neutral.
Inputs: query (required), k, as_of, intent, +1 more.
statsB
What the corpus contains: counts by source, kind and trust.
Inputs: no named parameters.
timelineA
How belief about a subject changed over time.
Matching claims in assertion order, showing which superseded which. Use when an answer may have changed rather than merely being unknown.
Inputs: subject (required), limit.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool targets a clearly distinct concern: recall for answering questions, timeline for temporal evolution, conflicts for unresolved contradictions, and stats for corpus overview. There is negligible overlap because recall returns ranked claims, while timeline explains changes and conflicts lists unresolved issues.
The names are short, lowercase, and readable, but there is no uniform convention: 'recall' is a verb, while 'timeline', 'conflicts', and 'stats' are nouns. This mixed style is still clear and memorable, so it is not chaotic, but it lacks the consistent verb-noun pattern seen in highly coherent tool sets.
Four tools is an appropriate scale for a read-only claim graph query server. Each tool covers a distinct mode of interaction without redundancy, and the count is well within the ideal range for a focused server.
The core workflows are covered: querying (recall), temporal analysis (timeline), contradiction handling (conflicts), and statistics (stats). Minor gaps exist, such as no way to list all claims or fetch a claim by ID directly, but these are workarounds that do not severely hinder the intended use.
Maintenance
Related MCP Connectors
Provenance-first database for teams and agents: every value carries sources, rules and coverage.
Query your own knowledge graph from any MCP client. The graph lives in a database you own.
Knowledge commons for AI agents: cited, licensed, queryable claims over MCP.
Read-only Kyntrax definitions, claims, evidence status and public provenance.
Related MCP Servers
- FlicenseNot gradedqualityCmaintenanceRead-only MCP server for querying an evidence-aware knowledge vault with temporal and provenance-aware data, supporting agent memory and semantic graph projections.-
- AlicenseNot gradedqualityCmaintenanceProvides an MCP server for storing and querying knowledge as verifiable claims, enforcing evidence-backed assertions with exact quotes and refusing paraphrases or unsupported relations.Apache 2.0
- AlicenseBqualityBmaintenanceEnables software agents to maintain a signed, evidence-aware local ledger of claims, with policy-gated proposals, conflict queries, and constitution-pinned access through an MCP stdio server.73Apache 2.0
- AlicenseNot gradedqualityCmaintenanceEnables agents to query a single-source ontology of types, entities, statements, evidence, and controlled vocabulary, including bitemporal "as of" and "as recorded" fact lookups. Supports writing by letting agents propose candidate statements while keeping approval, review transitions, and authoritative reasoning restricted to approved facts.MIT