Pinecone — Full Profile ( 2026-09)

Headquartered in San Francisco, California, Pinecone Systems Inc. was founded in 2019 by Edo Liberty, former Amazon AI executive, alongside co-founders Bob Wyman and Ram Sriharsha. The company has secured $138M in total venture capital backing from investors including a16z, Menlo Ventures, IVP and Tiger Global, with its last publicly reported valuation reaching $750M following its 2023 Series B financing round. Pinecone operates with a VC-independent market posture aligned with n8n, Databox, Apify and Dify, while maintaining a strictly managed-only product model. Unlike peer platforms, it offers no core self-hosted deployment capability and exclusively supports bring-your-own-cloud (BYOC) architecture. Pinecone delivers an official first-party GA developer MCP solution, published as the Apache-2.0 licensed open-source repositoriespinecone-database/mcp and pinecone-io/pinecone-mcp. As of February 2026, the project holds approximately 59 GitHub stars, a significantly lower community footprint compared to vector database peers Qdrant (1.3K stars) and Chroma (515 stars), while retaining full official first-party maintenance. Pinecone’s MCP implementation supports only stdio-based execution via npx, alongside per-assistant deployable options including managed remote endpoints and self-hosted Docker containers; notably, it does not provide a persistent hosted domain endpoint such as mcp.pinecone.io, differentiating it from hosted URL-based MCP services from Tavily and Apify. Functionally, its MCP toolset is strictly limited to Pinecone’s native integrated-inference indexes, with no support for external embedding index management. Ecosystem-wise, Pinecone joins the mainstream stable first-party GA server cluster consisting of Klaviyo, Feedly, Slack, Databox, Tavily, Apify and Dify, while carrying unique technical and operational differentiation within the cohort.

Website

Field	Value
Official Website	https://www.pinecone.io · docs: docs.pinecone.io · MCP guide: docs.pinecone.io/guides/operations/mcp-server · pricing: pinecone.io/pricing · GitHub MCP: github.com/pinecone-io/pinecone-mcp (Apache-2.0, ~59★ Feb 2026) · npm: @pinecone-database/mcp · console: app.pinecone.io
Company	Pinecone Systems Inc., San Francisco; founded 2019 Liberty+Wyman+Sriharsha; $138M a16z/Menlo/IVP/Tiger, ~$ 750M val 2023; vc-independent-leaning managed-only (no AGPL core, BYOC deploys control plane into customer VPC but storage plane still Pinecone-managed); SOC 2 Type II + ISO 27001 + HIPAA BAA + GDPR; SSO/SAML/RBAC/audit on Standard+; 100B+ queries served, 2.6B queries reference 2025
Status	Active, Serverless GA (2024) on AWS/GCP/Azure, Nexus knowledge-engine preview 2026, MCP developer-server GA, Assistant-MCP remote+Docker
Category	ai-productivity-data
Pipeline Stage	none

Title

Field	Value
Product	Pinecone (managed serverless vector database + Inference + Assistant + Nexus)
Descriptive Title	Pinecone — Fully-Managed Serverless Vector Database for RAG & Agent Memory: write-ack <100ms, auto-index rebalance, ANN query p99 flat at billion-scale, dense+sparse+full-text hybrid, metadata filter, integrated inference (embed+rerank inside Pinecone) + external embedding support; REST gRPC Python/TS/Go/Java SDK Bearer key; first-party MCP developer-server GA `@pinecone-database/mcp` stdio (tools: search-docs/list-indexes/describe-index/describe-index-stats/create-index-for-model/upsert-records/search-records/cascading-search/rerank-documents, integrated-inference indexes only) + Assistant MCP (per-assistant remote endpoint or Docker) for agent context retrieval; pricing = storage GB + read units (1 RU per GB namespace, 0.25 floor) + write units (1 WU per 1KB upsert, 5 floor) + egress, Starter $0 (2GB/2M WU/1M RU) → Builder$ 20 flat (10GB) → Standard $50 min (pay-as-you-go$ 0.33/GB $4/M WU$ 16/M RU) → Enterprise $500 min (BYOC/audit/SCIM); 2019 SF Liberty$ 138M a16z MCP-firstparty-GA-stdio-dev-plus-assistant-remote distinct-from-tavily-apify-hosted-url
One-line Positioning	Agent RAG pipeline upserts 1M chunks → Pinecone serverless indexes with integrated embedding → Claude mounts `@pinecone-database/mcp` stdio, calls search-records+cascading-search+rerank-documents → billing is storage+RU/WU+egress not per-seat (reverse Motion) not per-call (reverse Qualtrics) not CU (reverse Apify) not credit (reverse Tavily) not image-count (reverse QuickChart); ownership vc-backed-private managed-only (same macro-bucket n8n/Databox/Apify/Dify, distinct from Tavily acquired-kept / Zotero-CDS nonprofit / QuickChart bootstrap / Semantic Scholar AI2).

Features

Field	Value
Vector Core	Serverless (2024+) storage-compute separation, auto-scale from 0 to billions, ANN Rust index, p99 flat, dense (cosine/dot/euclidean) + sparse BM25 + full-text hybrid, metadata filter (numeric/string/array), namespaces, 1–3072 dims, integrated inference (pinecone-embed-v0, multilingual-e5, bge, Cohere/OpenAI passthrough) or BYO vectors
Write Path	Upsert ack <100ms, searchable in seconds, batch upsert, async import from S3/GCS/Azure Blob ($0.25/GB), record-level TTL, s-cache (prompt-cache layer shown in console)
Query Path	query/top_k/fetch/list, metadata filter pushdown, hybrid rerank (pinecone-rerank-v0, bge-reranker-v2-m3), cascading cross-index search dedup+rerank, namespaced isolation for multi-tenant
Inference	Pinecone Inference API (embed text/images, rerank) metered separately from DB; Assistant API (chat over uploaded files, citations)
Nexus	2026 preview “knowledge engine for agents” — compiles enterprise data into governed knowledge once, serves single-query cited answers, 90% fewer tokens, 30x faster than agentic RAG per Sierra τ-bench
MCP (first-party GA, two servers)	Developer MCP `@pinecone-database/mcp` Apache-2.0 stdio-only (npx, Node 18+, PINECONE_API_KEY env), 9 tools, integrated-inference indexes only (external-embedding indexes NOT supported via MCP), no hosted mcp.pinecone.io URL, no OAuth (API key only) → first-party-GA-server cluster but stdio-app-scoped like Dify-per-deployment not Tavily-hosted; Assistant MCP = per-assistant managed remote endpoint OR self-hosted Docker extending assistant context to Claude/Cursor; Claude Code plugin (Feb 2026) `/pinecone:quickstart` slash commands; Cursor/Gemini-CLI plugins
API	REST + gRPC, Bearer key, Admin API (service-account→project→scoped key programmatic provisioning), rate limits per project, 429 backoff; SDK Py/TS/Go/Java
Enterprise	SSO/SAML RBAC (Standard+), Audit logs/SCIM/CMEK/PrivateLink/Service Accounts (Enterprise), BYOC (Pinecone control plane in customer VPC, outbound-only, zero inbound SSH), 99.95% SLA Enterprise, HIPAA BAA

Pricing (2026-09, four-meter: storage GB + read units + write units + egress, plan floor min)

Field	Value
Starter	$0 — 2 GB storage, 2M WU/mo, 1M RU/mo, 5 indexes (dense/sparse/full-text), community Discord, AWS us-east-1 only, no SLA
Builder	$20/mo flat — 10 GB storage, 5M WU, 2M RU, multiple projects/users, Datadog/Prometheus monitoring, no overage billing (quota-blocked)
Standard	$50/mo minimum applied to usage (3-wk trial$ 300 credit) — pay-as-you-go: storage $0.33/GB, write units $ 4/M, read units $16/M, egress$ 0.10/GB after 100GB incl; DRN/backup/restore/RBAC/SSO/HIPAA-add-on
Enterprise	$ $500/mo minimum** — RU$ 24-27/M, WU $6-6.75/M, 99.95% SLA, BYOC, PrivateLink, CMEK, audit, SCIM, Pro support 24/7
BYOC	Quote-only — Pinecone in customer AWS/GCP/Azure VPC, outbound-only ops
Read Unit Math	1 RU per 1 GB of targeted namespace, 0.25 RU floor per query (top_k doesn’t change RU; a 10GB ns query = 10 RU regardless of top_k=1 or 100) — bill driven by namespace size not query count
Write Unit Math	1 WU per 1 KB upsert request, 5 WU floor per request (768-dim record ~3.2KB = 5 WU; batch 100 = ~357 WU)
Type	storage-ru-wu-egress-four-meter-plan-floor; reverse Motion per-seat, reverse Qualtrics per-interaction, reverse Tavily endpoint-credit, reverse Databox AI-credit, reverse QuickChart image-count, reverse Apify CU (both usage but RU/WU≠CU), reverse Gmail 80M-unit, reverse Semantic Scholar 1-RPS, reverse Zotero storage-quota, reverse Dify seat+credit
Gotcha	$50 Standard floor is floor-not-credit (a$ 18 month = billed $50); RU scales with namespace GB not query volume (fat namespace = silent 10x); MCP only talks to integrated-inference indexes (your OpenAI-embedded index invisible to MCP tools); egress 4th meter since mid-2026

Reviews

Field	Value
G2	4.4/5 ~240-260 reviews (devtool-B2B SaaS pattern, real claimed profile) → `G2 4.4/~250` aligns Klaviyo/Databox/Apify/Dify vc-SaaS G2-rated, thinner than Klaviyo 580, solid-er than Tavily 2-review; praise serverless zero-ops/sub-100ms, gripe RU-opacity/namespace-bill-shock
Capterra	4.6/5 ~30-50 reviews → `Capterra 4.6/~40` directional-solid
Trustpilot	No B2C page (B2B infra) → `no-rated-consumer-profile` excluded (aligns FFmpeg/QuickChart/Semantic Scholar/Zotero/Apify)
Dev / MCP sentiment	pinecone-mcp 59★ (Feb 2026) far below Qdrant 1.3K/Chroma 515; AIMultiple/Context.dev 2026 “category-defining managed vector DB, MCP late and stdio-only integrated-inference-only”; praise Nexus direction/cascading-search uniqueness, gripe MCP narrow scope (no external-embedding index, no hosted URL, 59★ shows low agent-adoption yet)
Praise	Serverless auto-scale zero-capacity-planning, sub-100ms ack, hybrid dense+sparse+full-text, integrated inference removes embedding orchestration, cascading-search cross-index dedup+rerank unique among vector MCPs, BYOC for regulated, Nexus 30x claim
Gripes	RU bill shock (namespace-size not query-count), $50/$ 500 floor burns small teams, MCP stdio-only no hosted URL, MCP integrated-inference-only blind spot for BYO-embedding users, no self-host OSS core (Qdrant/Weaviate escape), 59★ MCP shows agent-ecosystem lag vs Qdrant

Review Sources Breakdown

Field	Value
G2.com	4.4/5 ~250 claimed devtool-B2B → `G2-solid-vc-saas-pattern`
Capterra	4.6/5 ~40 → `Capterra-solid`
Trustpilot	none → `no-rated-consumer-profile` excluded
GitHub pinecone-io/pinecone-mcp 59★ + Qdrant 1.3K comparison	dev-sentiment directional (MCP adoption lag)
AIMultiple/Context.dev	directional 4.4-norm
Conclusion	Prints “G2 4.4/~250 + Capterra 4.6/~40 devtool-B2B real (not Tavily 2-review, not Zotero/QuickChart absent), GitHub MCP 59★ (low vs Qdrant 1.3K, integrated-inference-only scope), dev consensus serverless-zero-ops-RU-opaque-MCP-late-stdio” — aligns vc-independent SaaS G2-rated pattern.

Access

Field	Value
Web App	app.pinecone.io — index console, s-cache view, namespaces, metrics, API keys, Nexus preview
API	REST + gRPC Bearer key, Admin API (client_cred→token→project→scoped key), SDK Py/TS/Go/Java, LangChain `PineconeVectorStore`, LlamaIndex `PineconeVectorStore`, Haystack, n8n HTTP node, Airbyte connector
MCP	First-party GA, two forms — (1) Developer MCP: `npx -y @pinecone-database/mcp` stdio, PINECONE_API_KEY env, Node 18+, 9 tools, integrated-inference indexes only, no hosted mcp.pinecone.io, no OAuth; (2) Assistant MCP: per-assistant managed remote endpoint OR `docker run pinecone/assistant-mcp` self-hosted, gives Claude/Cursor context from assistant files; (3) Claude Code plugin Feb 2026 `/pinecone:*` slash; Cursor/Gemini-CLI plugins — joins first-party-GA-server cluster but stdio-app-scoped (like Dify-per-deployment, unlike Tavily/Apify hosted URL)
Login	app.pinecone.io email/GitHub/Google → API Keys tab → scoped key
Self-host	No OSS core (contrast Weaviate/Qdrant/Chroma); BYOC deploys Pinecone control plane into customer VPC but storage/index still Pinecone-managed code, outbound-only; Assistant MCP Docker is self-hostable but talks to Pinecone cloud

Best For

Field	Value
Audience 1	RAG teams wanting zero-ops vector layer (no Redis+Qdrant+embed-server to run), BYO OpenAI/Cohere embeddings or use integrated inference
Audience 2	Agent builders needing long-term memory store with metadata filter + hybrid search, Claude mounts `@pinecone-database/mcp` for cascading-search+rerank
Audience 3	Multi-tenant SaaS isolating per-customer namespaces (1 RU/GB makes small-tenant cheap, large-tenant predictable)
Audience 4	Enterprises needing SOC2+HIPAA+BYOC+PrivateLink for regulated retrieval (health/financial)
Audience 5	Teams adopting Pinecone Nexus (2026) to replace agentic-retrieve-loop with compiled knowledge (90% token cut claim)
Audience 6	E-commerce / support semantic search at 100M+ doc scale with dedicated read nodes

Not For

Field	Value
Exclude 1	Teams wanting OSS self-host on own Postgres/Redis — use Qdrant/Weaviate/Chroma/pgvector (Pinecone BYOC ≠ self-host)
Exclude 2	BYO-embedding users expecting MCP to touch those indexes — Pinecone MCP integrated-inference-only, external-embedding indexes invisible to MCP tools
Exclude 3	Tiny projects allergic to $50 floor — Starter 2GB/2M WU is real free but Standard jumps to$ 50 min even at $5 usage
Exclude 4	Teams assuming RU = per-query — bill scales with namespace GB, fat single-namespace = 10x surprise (shape namespaces or eat it)
Exclude 5	Buyers needing hosted mcp.pinecone.io OAuth URL like Tavily/Apify — Pinecone MCP is stdio-only key-auth
Exclude 6	Pure keyword/search-engine replacement — Pinecone is ANN not inverted-index (use Elastic/OpenSearch for that)

Competitors

Field	Value
Weaviate	OSS Apache-2.0 + managed, hybrid search mature, 25K★+, multi-tenant, GraphQL, worse zero-ops than Pinecone but self-hostable, MCP community-only
Qdrant	Rust OSS MIT + Qdrant Cloud, 13K★+, MCP server 1.3K★ (beats Pinecone 59★), payload filter strong, cheaper 100M-vector ~$33 self-host, less enterprise BYOC polish
Chroma	OSS lightweight (515★ MCP), dev-first, not billion-scale production, no BYOC enterprise
pgvector	Postgres extension, $0 if already on PG, 1536-dim HNSW, no separate infra, weak at 100M+ scale, no integrated inference
MongoDB Atlas Vector	HNSW in Atlas, Voyage-AI embed add-on, doc+vector same collection, M0 free, not purpose-built ANN
Milvus / Zilliz Cloud	OSS Milvus + Zilliz managed, GPU Cardinal engine, billion-scale, heavier ops than Pinecone
Redis Stack / Vald / Marqo	secondary alternatives