FreshContext

This package is no longer developed. 0.5.3 is the last release of freshcontext-mcp. It and every earlier release stay available under the MIT License, as is; there will be no further feature releases. For FreshContext services, see https://freshcontext.dev. Security reports: see SECURITY.md.

I asked Claude to help me find a job. It gave me a list of openings. I applied to three of them. Two didn't exist anymore. One had been closed for two years.

Claude had no idea. It presented everything with the same confidence.

That's the problem freshcontext fixes.

This repository is the integrated FreshContext Core/MCP package.

Category: context integrity infrastructure. FreshContext sits between context acquisition and agent action. Its job is to decide whether information entering an AI workflow is still fresh, attributable and coherent enough for the system to rely on. Core is the reusable engine that scores, ranks, explains and turns candidate context into decision-ready context, with signed verdicts recorded in a verifiable ledger. MCP is the first live host interface over that engine — one interface over the methodology, not the product itself.

Live demo: api.freshcontext.dev/demo — same model, same query, two completely different answers. Only the temporal layer changed.

Integrate it into an existing stack: freshcontext.dev/integration — start with one bounded RAG, agent, retrieval, or governance workflow and objective acceptance criteria.

The problem

Large language models retrieve web data semantically. Cosine similarity finds the documents that match a query best — but cosine doesn't know when a document was written.

So a 2022 blog post and a 2026 paper can score nearly identically. The model gets a context window full of stale documents and faithfully summarizes 2022 advice for a 2026 question.

That's not hallucination. That's correct summarization of corrupted retrieval.

Most RAG pipelines rank context correctly semantically but incorrectly temporally.

The layer

FreshContext is context integrity infrastructure for AI agents and retrieval systems. It sits between retrieval and reasoning:

candidate context
  -> FreshContext Core
  -> decision-ready context
  -> model / agent / app

FreshContext evaluates freshness, source profile, confidence, utility, provenance material, and failure honesty before context reaches the LLM. The temporal core uses Decay-Adjusted Relevancy:

R_t = R_0 · e^(−λt)

• R_0 — base semantic relevancy (whatever your retriever already gives you)
• λ — source-specific decay constant (HN ≈14h half-life, blogs ≈29d, academic papers ≈1.6y)
• t — hours elapsed since publication
• R_t — decay-adjusted relevancy at query time

That's the core correction. No model swap. No re-embedding. No re-indexing. The layer drops onto whatever retrieval pipeline you already have.

The layer is the product. The named adapters shipped with this repo demonstrate compatibility across different source classes. The DAR engine, the freshness envelope, Source Profiles, and the FreshContext Specification are the moat.

The standard

Every FreshContext-compatible response wraps content in a structured envelope:

[FRESHCONTEXT]
Source: https://github.com/owner/repo
Published: 2024-11-03
Retrieved: 2026-03-05T09:19:00Z
Confidence: high
---
... content ...
[/FRESHCONTEXT]

When it was retrieved. Where it came from. How confident we are the date is accurate.

The FreshContext Specification v1.2 is published as an open standard under MIT licence. Any tool, agent, or system that wraps retrieved data in this envelope is FreshContext-compatible. → Read the spec · Read the methodology

Architecture boundary

FreshContext Core is the reusable center of the current integrated package. It owns signal normalization, freshness scoring, Source Profiles, decision output, envelope formatting, failure guards, shared types, rank/explain primitives, and the context-conditioned utility primitive.

MCP is the primary reference/interface implementation over Core. Claude Desktop is supported, but not required. The MCP tool surface exposes named reference adapters and a live interface for using the system.

The production Cloudflare Worker now uses Core-backed envelope generation. Worker-specific concerns remain outside Core: MCP transport, runtime guards, KV cache policy, cache metadata injection, JSON parse/replace cache helpers, D1 feeds, cron, rate limiting, and Store/feed scoring/provenance.

See Architecture for the layer boundaries, the package surface, and what is not yet separated.

Core import path

FreshContext Core is also available directly from the current MCP package:

import {
  evaluateSignals,
  interpretEvaluations,
  getSourceProfile,
  normalizeSignal,
  calculateHaPriV2,
} from "freshcontext-mcp/core";

This is a Core subpath export inside freshcontext-mcp, not a standalone freshcontext-core package yet. The root package and freshcontext-mcp binary remain the MCP reference host.

Primary MCP interface

The clearest MCP path is evaluate_context.

It accepts candidate context from any retriever, agent, database, local script, note parser, or adapter output:

{
  "profile": "academic_research",
  "intent": "citation_check",
  "signals": [
    {
      "title": "Example source",
      "content": "Candidate context text...",
      "source": "https://example.com/source",
      "source_type": "arxiv",
      "published_at": "2026-05-24T12:00:00.000Z",
      "retrieved_at": "2026-05-24T13:00:00.000Z",
      "semantic_score": 0.92
    }
  ]
}

FreshContext returns decision-first output:

• Decision
• Meaning
• Action
• Warnings
• Source
• Freshness
• Rank score
• Utility
• Confidence
• Why

Structured results also include a readable object for humans:

{
  "decision": "cite_as_primary",
  "label": "Cite as primary",
  "readable": {
    "label": "Primary source",
    "summary": "This source is strong enough to use as main evidence.",
    "why": [
      "Strong semantic match and current freshness for arxiv.",
      "source profile academic_research uses lenient date policy",
      "intent profile citation_check selected"
    ],
    "action": "Use this as main evidence while preserving citation and provenance.",
    "warnings": [
      "FreshContext judges citation readiness and context usefulness; it does not certify truth."
    ]
  }
}

The readable object translates Core decisions into user-facing language. It does not change ranking, decision labels, utility scoring, or source intake. Utility helps explain usefulness for the current question; it remains explanatory and does not control default decision labels or ranking.

FreshContext does not certify truth. It records why context was used, supported, questioned, refreshed, watched, or excluded before it reaches a model.

evaluate_context does not fetch URLs, crawl, scrape, browse, read folders, or call adapters. It only evaluates candidate context the caller provides.

Current boundary: evaluate_context ships in the npm/local stdio MCP server. The hosted Cloudflare Worker MCP endpoint is a separate deployment surface and is verified independently — check /v1/health for its live version and tool count rather than assuming parity with the package. The Worker remains a separate deployment surface, so future package interfaces should be re-verified remotely before being claimed live.

Network Boundary

FreshContext's primary evaluate_context path does not fetch, crawl, scrape, browse, read folders, or call adapters. The MCP package also includes read-only reference adapters that use network access only when those adapter tools are invoked. Supply-chain scanners may therefore report package network access; that applies to the optional adapter surface, not to caller-provided context evaluation.

Advanced Worker/feed surface

Beyond the per-call Core/MCP paths, the production Worker deployment exposes a continuous, decay-scored, deduplicated feed. This is an advanced deployment surface, not the required way to use FreshContext Core:

GET /v1/intel/feed/:profile_id?limit=20&min_rt=0

Every signal is stamped with base_score, rt_score, entropy_level (low / stable / high), ha_pri_sig (Ha-Pri v1 SHA-256 provenance reference), semantic_fingerprint (cross-adapter dedup), and published_at. Ready for direct LLM or agent consumption — no synthesis required.

Production endpoint: https://api.freshcontext.dev

Reference adapters

The repo ships named reference adapters that demonstrate how different source classes can become FreshContext-compatible. Each adapter keeps its own name because it represents a source boundary; the adapter count is operational proof, not the product headline.

Intelligence

Competitive research

Market data

Composites — multiple sources, one call

Official, regulatory, and procurement sources

Quick start

For Claude Desktop, Codex, npx, global npm, and source-checkout setup, see the concise client setup guide.

Cloud (no install)

Add to your Claude Desktop config and restart:

Mac: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "freshcontext": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://api.freshcontext.dev/mcp"]
    }
  }
}

Restart Claude. Done.

Prefer a guided setup? Visit freshcontext.dev — 3 steps, no terminal.

Local (full Playwright)

Requires: Node.js 20+ (nodejs.org)

git clone https://github.com/PrinceGabriel-lgtm/freshcontext-mcp
cd freshcontext-mcp
npm install
npx playwright install chromium
npm run build

Add to Claude Desktop config:

Mac:

{
  "mcpServers": {
    "freshcontext": {
      "command": "node",
      "args": ["/Users/YOUR_USERNAME/path/to/freshcontext-mcp/dist/server.js"]
    }
  }
}

Windows:

{
  "mcpServers": {
    "freshcontext": {
      "command": "node",
      "args": ["C:\\Users\\YOUR_USERNAME\\path\\to\\freshcontext-mcp\\dist\\server.js"]
    }
  }
}

Mac troubleshooting

"command not found: node" — Use the full path:

which node  # copy this output, replace "node" in config

Config file doesn't exist:

mkdir -p ~/Library/Application\ Support/Claude
touch ~/Library/Application\ Support/Claude/claude_desktop_config.json

Usage examples

The npm run demo:* commands below are source-checkout workflows for contributors and evaluators using a cloned repository. The published npm package is the MCP server/runtime package and does not include repo-only source examples or tests.

From an installed npm package, the supported runtime entrypoints are npm start and the freshcontext-mcp binary. Repo-only scripts such as tests, demos, smoke checks, and trust scans print a source-checkout notice when their source files are not present.

The Apify Actor entrypoint remains available in the source checkout for separate actor packaging, but it is intentionally not part of the published MCP npm runtime package.

Release trust gate

Run the local release gate before a release, package review, demo, or PR review:

npm run trust:gate

The gate runs the Trust Scanner with repo-map reporting, npm package-boundary inspection, deterministic claim checks, and --fail-on fail. It is local-only, does not publish or deploy, does not send telemetry, and does not replace dedicated security scanners.

Generate review reports when you need a shareable summary:

npm run trust:report
npm run trust:report:json

To write a Markdown report file explicitly:

npm run trust:report -- --output TRUST_SCAN_REPORT.md

Bring your own source list

FreshContext can evaluate candidate context you provide as a local JSON file:

npm run demo:evaluate:file

To pass a different file:

npm run demo:evaluate:file -- path/to/sources.json

Included examples:

npm run demo:evaluate:file -- examples/sources.academic.example.json
npm run demo:evaluate:file -- examples/sources.jobs.example.json

Minimal shape:

{
  "profile": "academic_research",
  "intent": "citation_check",
  "signals": [
    {
      "title": "...",
      "content": "...",
      "source": "...",
      "source_type": "arxiv",
      "published_at": "...",
      "retrieved_at": "...",
      "semantic_score": 0.92
    }
  ]
}

This local demo does not fetch URLs, crawl, or read folders. It evaluates candidate context you provide and returns decision-first output: Decision, Meaning, Action, Warnings, and supporting metrics.

In an MCP client, use evaluate_context when you already have candidate context from another retriever, database, agent, or script:

Use evaluate_context with profile "academic_research", intent "citation_check", and these candidate signals: [...]

Use the named reference adapters when you want FreshContext's current MCP package to fetch public source examples for you.

Should I build this idea?

Use extract_idea_landscape with idea "procurement intelligence saas"

Returns funding signal, pain signal, crowding signal, market signal, ecosystem signal, and launch signal — all timestamped.

Full company intelligence in one call:

Use extract_company_landscape with company "Palantir" and ticker "PLTR"

SEC filings + federal contracts + global news + changelog + market data.

Did that company just disclose something material?

Use extract_sec_filings with url "Palantir Technologies"

8-K filings are legally mandated within 4 business days of any material event — CEO change, acquisition, breach, major contract.

Is this dependency still actively maintained?

Use extract_changelog with url "https://github.com/org/repo"

Returns the last 8 releases with exact dates. If the last release was 18 months ago, you'll know before you pin the version.

Deployment & infrastructure

The reference implementation runs on Cloudflare's global edge:

• D1 database — 18 watched queries running on 6-hour cron with relevancy scoring
• KV-backed rate limiting — 60 req/min per IP across all edge nodes
• Defensive valves — clock-skew rejection (5min tolerance), hard floor at R_t<5, lazy decay at read time
• Provenance — feed signals still carry legacy Ha-Pri v1 SHA-256 provenance references; separately, ledger-backed context verdicts are signed with Ed25519 V4 and independently verifiable
• Schema migrations — promise-gated, idempotent, run on first request after deploy

Production: https://api.freshcontext.dev

Deployment modes

The engine is deliberately separable from the interface it is reached through. The same Core runs in each of these without a rewrite:

Only the MCP and standalone modes are exercised in production today. The others are integration seams the architecture already supports, not shipped configurations.

Roadmap

Split three ways so that genuine engineering risk is never filed as optionality. Nothing outside Production core is a live product claim.

Production core — built, running, testable

• FreshContext Specification v1.2 published (MIT, open standard)
• DAR engine with source-specific lambda constants
• Ha-Pri v1 provenance signatures on stored signals
• Ha-Pri v2 Core helper and deterministic golden vectors
• Public /v1/verify endpoint — ledger-backed verdict verification, answering for both the legacy HMAC path and Ed25519, and reporting which was used via verification_method
• Generic MCP evaluate_context tool for caller-provided candidate context
• Core-backed envelope generation shared by npm/MCP and the Cloudflare Worker
• Semantic deduplication via fingerprinting
• Named reference adapters across intelligence, competitive research, market data, and composites
• Cloudflare Workers deployment — global edge, KV cache, atomic rate limiting
• Live before/after demo at /demo
• METHODOLOGY.md — methodology and engineering documentation
• Published on npm and listed for MCP usage; Apify/feed assets separated from the MCP runtime package
• Trusted release publishing workflow — manual workflow_dispatch only, OIDC-backed, provenance-enabled, and gated by version/verification checks. One explicit run publishes npm first, verifies it, then publishes the matching manifest to the official MCP Registry with GitHub OIDC
• Independently verifiable Ed25519 attestation (E-2). Every new verdict row in the ledger is signed FRESHCONTEXT_HA_PRI_V4 with Ed25519. A third party can verify a verdict with no FreshContext account, no API key and no call to FreshContext — using the key document the Worker publishes at /.well-known/freshcontext-signing-keys.json and either verifier shipped in the npm tarball: scripts/verify-offline.mjs (Node, standard library) or scripts/verify_offline.py (Python, no dependencies at all). Written up for the sceptic rather than the maintainer in VERIFYING.md.
• Signing key fc-2026-09-ceced1ab published and active. Keys are append-only, so a rotation never invalidates a verdict signed under a key that has since been retired.
• attestation-proof.yml — obtains a live verdict, verifies it with both shipped verifiers, runs tampered-payload and tampered-signature negative controls, and confirms the stored ledger row is V4 rather than only the emitted response block. On demand and daily; every input it uses is public, so it needs no credentials to run.

In flight on the core, not an expansion surface:

• Ha-Pri v2 Worker/D1 production enforcement for stored signals — the feed rows, which still carry Ha-Pri v1 SHA-256 stamps. This is a separate path from the verdict ledger above: verdicts are V4/Ed25519 today, signals are not. Design document complete; hard tamper enforcement on the signals path is not live.

Expansion surfaces — deliberately open, not built

These are integration seams the architecture supports and the engine does not yet implement. Stated in future tense on purpose.

• Context safety harness. Policy enforcement before context reaches an agent: pass / warn / refresh / quarantine / block, with evidence attached to each decision. Today evaluate_context emits decisions and warnings; the enforcement state machine does not exist — quarantine and block are not implemented anywhere in the codebase.
• Enterprise control plane. Dashboard over source health, trust score, context drift and provenance lineage. The verdict ledger is the data contract this would read from; the UI is unbuilt.
• Observability telemetry. Historical integrity state, incidents, upstream degradation and remediation history.
• Autonomous remediation. Automatic refresh, source substitution and re-evaluation — closed-loop rather than detection-only.
• Vertical policy packs. Domain-specific integrity thresholds for regulated workflows.
• Webhook triggers — push high-entropy signals on threshold

Research frontier — exploration, not commitment

• GKG upgrade for extract_gdelt — tone scores, goldstein scale, event codes
• Contradiction detection across concurrent sources

Future work is organized in FreshContext Future Lanes. Roadmap items are not live product claims until implemented and validated.

Contributing

PRs welcome. The highest-value contributions improve the caller-provided context path, decision output, host integrations, and FreshContext-compatible signal quality. New reference adapters are useful when they preserve source boundaries and emit timestamped, failure-honest context — see src/adapters/ for examples and FRESHCONTEXT_SPEC.md for the compatibility contract.

If you're building something FreshContext-compatible, open an issue and we'll add you to the ecosystem list.

Trust and security

• LICENSE
• SECURITY.md
• NOTICE.md
• TRADEMARKS.md
• Dependency diligence notes
• Release integrity notes
• Release notes

License

MIT

Built by Immanuel Gabriel — Namibia 🇳🇦 "The work isn't gone. It's just waiting to be continued."

Also on: MCP Registry · npm