Live context for agents
Coalent turns your sources into dense, attributed facts — served at a fraction of RAG’s tokens and invalidated the instant a source changes.
Never stale.Always attributed.
pricing.md changedpro_plan: $29 $392 claims rebuilt 0.4spricing.md$29 $39fresh 0.4spip install coalent// apache-2.0 · zero-dependency core · built in the open
# Pricing
- Free — $0/mo
- Pro — $29/mo$39/mo
- Team — $99/mo
[pricing.md | modified today]
pro_plan: $29/mo
free_tier: $0/mo
Pro plan is $29/mo.
fresh · 0.4s
n=605 held-out · strict grading · third-party gold questions — method →
compresses the bytes of each call — per-request, stateless.
evolves an agent’s playbook — how to act, not what’s true.
remembers the person across sessions — who you’re talking to.
builds knowledge graphs over your corpus, upfront.
dense claims derived from your sources, invalidated by provenance the instant a source changes.
Coalent doesn’t compete for who-you-are or how-to-act — it complements the memory layers and owns the freshness slot: what your sources say right now, with context invalidation that is surgical — span-level provenance, not a TTL guess. That’s why its layer sits directly on your sources.
Sources
docs · tickets · APIs
Point the cache at wherever your knowledge lives. Units build lazily, on first read — no upfront graph tax.
Atomic claims
extract once · span-attributed
Each source becomes query-independent claims — every number kept, every claim pinned to the span it came from.
Budget-packed serving
read_path="pool"
Every read gets the globally-ranked fresh claims, packed to your token budget under per-source attribution headers.
Surgical invalidation
source Δ → its claims only
A changed source stales exactly the claims derived from it. They rebuild on the next read — nothing else is touched.
The same cache behind an MCP context server for Claude Code and Cursor, or as a LangChain cache over the vector store, embeddings, and chat model you already run. Nothing about how you built them changes.
- CLAUDE CODE
- CURSOR
- MCP
- LANGCHAIN
- ANY VECTOR DB
MCP — CLAUDE CODE · CURSOR · ANY CLIENT
LANGCHAIN — YOUR EXISTING STACK
// folder mode (--watch) is zero-setup with a measured cost: 0.46 vs 0.71 in factory mode — bring your own retriever for the full number
Dense.
Durable.
True.
We measured naive RAG’s own token-scaling curve on the same stream — so Coalent is compared against every operating point of the baseline, not one convenient k. Real news corpus, third-party gold questions, n=605 held-out, strict grading.
ACCURACY
CONTEXT TOKENS / READ
| ARM | ACCURACY | CTX TOKCONTEXT TOK / READ |
|---|---|---|
| naive k=4 | 0.582 | 590 |
| naive k=6 | 0.638 | 882 |
| naive k=9 | 0.711 | 1,311 |
| naive k=12 | 0.731 | 1,729 |
| Coalent pool | 0.731 | 981 |
The honest framing: we do not claim to beat naive RAG on accuracy — its own curve reaches the same number given more tokens. The claim is equal accuracy at 25–43% fewer context tokens, plus what no k can buy: stale RAG answers are masked the instant a source changes, and every fact carries its span.
GITHUB
★ 9
Apache-2.0, in the open — the source, the issues, and the eval code behind every number on this page.
Vectorlink-Labs/coalent →
PYPI
pip install coalent
Zero-dependency core. Extras are opt-in: [mcp] for the server, [openai] for folder mode.
coalent on PyPI →
MCP REGISTRY
coalent-mcp
The MCP server, one command from any client — validated byte-identical to the library.
MCP registry →
v0.6.1 aug 2026 · coalent-mcp + langchain-coalent · v0.6 jul 2026 · read_path="pool"
# dense, durable, true — live context for your agents
pip install coalent