Claims Mcp

io.github.dgonzap30Published by GitHub account dgonzap30v0.1.0Updated Oct 24, 2026

A claims memory for agents: assert beliefs, let them decay, surface contradictions.

VerifiedSTDIODesktop onlyKnowledge & MemoryAI & ML

Overview

AI-generated overview

A local claims memory for agents that stores beliefs about entities, decays them over time, and flags stale or contradictory claims.

What you can do
Stores claims about entities with a kind, volatility, confidence and state, and lets an assistant assert, search, confirm or retract them. Claims decay on a half-life set by volatility, so effective confidence falls until a claim is re-confirmed. Tools surface stale claims and contradictions such as duplicate priority ranks or conflicting status claims, and a resource exposes all claims as JSON.
When to use it
Useful when an assistant needs a small persistent memory of beliefs that should age, for example tracking priorities, statuses or relationships across sessions and noticing when they go stale or conflict. Best suited to single-user local workflows where the assistant supplies its own observations.
Requirements
Runs as a local stdio process on the user's machine, installed from the npm package claims-mcp and started with bunx. Requires Bun 1.1 or newer, since it uses bun:sqlite rather than plain Node. Data is kept in one local SQLite file at ~/.claims-mcp/claims.db, overridable with --db or the CLAIMS_MCP_DB environment variable. No network, accounts or API keys.
Before you install
It writes to a local SQLite database, so claims can be added, updated, confirmed or retired and persist between sessions. The database path can be changed with CLAIMS_MCP_DB or --db, so point it somewhere you intend. No credentials or network access are requested, and observations used for rank-drop checks are supplied by the caller rather than collected.

Installation

In SourceWeft

  1. Open Claims Mcp in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

claims-mcp

A small memory for agents that treats what it knows as claims that age. You assert a claim about an entity ("acme is priority #1", "the launch is on track"), each claim decays on a half-life set by how volatile it is, and the server tells you which claims have gone stale and which contradict each other.

It is a stdio MCP server plus a CLI over one local SQLite file. No network, no accounts.

Runtime: Bun. v0 uses bun:sqlite, so it runs on Bun ≥ 1.1, not plain Node. A Node build (node:sqlite) is a possible follow-up.

Model

FieldMeaning
entityWhat the claim is about (free text key).
kindpriority (may carry assertedRank), status, fact, relationship.
volatilitystable (half-life 365d), slow (30d), volatile (7d).
confidence0..1 at last confirmation. Effective confidence halves every half-life since lastConfirmed, floored at 0.05.
stateactive, stale, contradicted, retired. Only active claims are checked.

A claim's identity is (entity, kind, claim): asserting it again updates it rather than duplicating.

MCP tools

ToolDoes
claim_addAssert or update a claim.
claim_searchCase-insensitive substring search over claim text and entity.
claim_confirmRe-confirm now: resets decay, reactivates, optionally sets confidence.
claim_retractRetire a claim.
claims_contradictionsConflicts among active claims: duplicate-rank (two entities claim the same priority rank), status-conflict (several active status claims for one entity). Pass observed weekly ranks to also get rank-drop: an asserted-priority entity that every signal ranks worse than rank + 2 for 3 valid periods.
claims_decayedActive claims past a staleness threshold (default 0.5), stalest first.

Resource: claims://all (every claim as JSON).

observed is { "signals": [{ "<entity>": [rank_now, rank_1_period_ago, ...] }, ...], "validPeriods": [true, ...] }. You supply the observations (commit counts, time spent, anything you can rank); the server does not collect them.

Use

bash
bun install

Add to an MCP client (Claude Code shown):

bash
claude mcp add claims -- bun /path/to/claims-mcp/src/mcp.ts

The database lives at ~/.claims-mcp/claims.db; override with --db <path> or CLAIMS_MCP_DB.

CLI:

bash
bun src/cli.ts add acme priority slow 0.8 "acme comes first" --rank 1bun src/cli.ts search comesbun src/cli.ts stale 0.5bun src/cli.ts contradictionsbun src/cli.ts confirm <id>bun src/cli.ts retract <id>

Development

bash
bun test          # 19 tests, synthetic claims onlybun run typecheckbun run audit     # publish gate: home paths, private names, secret patterns

Set AUDIT_TERMS="name1|name2" to add your own private names to the audit.

Provenance

Not a straight extraction. The decay curve and the rank-drop rule are ported from a personal agent memory (rank-drop now takes caller-supplied observations); the duplicate-rank and status-conflict rules are new claim-vs-claim logic written for this package. It shares no code, data or storage with the original.

Author

Built by Diego González Zapiain at Lojik Labs. Listed on the official MCP Registry as io.github.dgonzap30/claims-mcp.

Licence

MIT

Source: README.md at commit 66c1df8

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.0LatestSep 20, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts