Convex

What is Convex?

Convex is a reactive backend platform that combines a database, server functions, and client libraries. Its queries are TypeScript code running alongside the database, and subscribed results react to changes in stored data.

The Convex realtime database uses a relational data model while storing JSON-like documents. Schemas are optional, documents can reference documents in other tables, and application code reads and writes through a JavaScript API rather than SQL.

Convex also provides managed AI model access through AI Gateway. This is a service called from backend actions, with provider credentials held by Convex and short-lived access tokens scoped to a deployment.

Core features

Queries, mutations, and actions

Function types define different execution boundaries:

• Queries read data, are cached, and support reactive subscriptions.
• Mutations write data and execute as database transactions.
• Actions call external APIs. They are not transactional or cached.
• HTTP actions expose webhook or custom-client endpoints.

Managed model access

AI Gateway exposes routes for text generation, embeddings, decisions, images, videos, speech generation, and transcription. Decisions, Voice, and image/video generation are alpha; their request/response formats may change.

Guide

Connecting a React interface

An existing Convex backend connects to a React interface through these primitives:

1. Create a ConvexReactClient for the deployment and pass it to a ConvexProvider surrounding the component tree.
2. Read backend data with useQuery and the generated api object. The development command generates that object in convex/_generated/api.js, with JavaScript autocompletion and end-to-end TypeScript types.
3. Change data with useMutation. The hook returns an asynchronous function; calling that function with its arguments performs the mutation.

Convex use cases

• Live chat interfaces: Convex provides a tutorial for a realtime chat application with persistent, live-updating messages.
• Semantic retrieval and recommendations: Vector search finds documents similar to a supplied vector. Its documented uses include context for LLM applications and related-content recommendations.
• Persistent agents: The Agent component persists conversation history across sessions and separates long-running AI work from a reactive UI. It supports hybrid vector/text search over that history and RAG-based prompt augmentation from other sources through the RAG component or a custom implementation.
• Delayed and recurring work: Scheduled functions run once later, while cron jobs recur. Examples include welcome emails and account reconciliation; Workpool and Workflow components add queues or durable flows with retries and delays.

Who is it for

• JavaScript and TypeScript application developers: The TypeScript backend also accepts JavaScript; functions are exposed through client-library APIs, fitting teams that own both interface and server code.
• Teams seeking hosted operation: Convex positions its hosted product as more hands-off; self-hosting runs the current backend code on the team's own servers.

Platforms

• React clients: ConvexReactClient uses a WebSocket for pushed results without polling and reconnects after connectivity drops.
• Frontend hosting: The frontend can use Vercel, Netlify, or custom hosting, separately from the backend deployment.
• Authentication: Connections/calls use OpenID Connect JWTs. Convex Auth supports client-side React and React Native signup/sign-in flows, but remains beta with fewer features than third-party integrations.
• Data regions: Deployments can use Northern Virginia, Ireland, Sydney, or Canada Central. Non-US usage has a 30% surcharge (1.3 times standard rates).

All infrastructure powering a deployment is hosted in its selected region. Changing region requires a new project or deployment and data export/import; an existing deployment cannot move in place.

Pricing

Prices were captured on November 2, 2026, in USD at the displayed US East baseline. Included resources are shared across regions.

Plans and developer access

Business lists a $2,500 monthly minimum; deployment-class and infrastructure unit rates are published separately. Enterprise lists annual commitment discounts without a numerical discount or annual total.

Free, Starter, and Professional resources

Each cell lists an included amount and an additional-use rate. In the first column, paid overages apply to Starter; Free has fixed allowances. Storage is billed per GB/month, while other resources have their stated meters.

Serverless query/mutation compute is free. Calls include client calls, schedules, subscription updates, and file accesses. File storage includes backups; egress includes function file access and backup generation/restoration. Text/vector indexes and queries share their respective search meters.

Usage is summed across the team's projects, not allocated separately per project or developer. Extended Free-plan overuse may cause HTTP errors, but Convex says successfully written mutation records are never deleted for exceeding resources. The FAQ names Starter and also uses “Starter Plus” for continued traffic. No separately priced tier is inferred from these labels.

Business and Enterprise deployment rates

The US East monthly-rate matrix lists four classes; active deployments are prorated daily. Other units appear in each row.

These infrastructure meters apply alongside the deployment class:

A dash preserves the published value without implying a zero price or included entitlement.

AI charges and spending controls

AI Gateway requires a paid team and supports cloud production/development or project-linked local deployments. Free and self-hosted deployments cannot use it; self-hosted actions can call providers with their own keys.

Gateway costs depend on model, timing, cached tokens, image/video resolution and duration, or voice characters/seconds/tokens. They appear on the Convex invoice; cancelling a submitted video request may still incur generation cost.

Cloud deployments support daily/monthly AI Gateway disable thresholds that disable the whole deployment, not just model requests; non-development deployments also support warning thresholds. Local project-connected requests share the team's spending limit; in-flight requests or already-issued short-lived credentials may finish after it is reached.

Renewal and refunds

Fees are noncancelable/nonrefundable except express contractual exceptions. Order Forms control monthly/annual cycles; self-service overages are automatic, with interim charges possible at an invoice threshold.

Order Form subscriptions auto-renew for an equal term unless either party gives nonrenewal notice at least 60 days before expiry. If the customer elects not to auto-renew, the subscription converts to month-to-month at current terms and pricing; either party may terminate it with 30 days' prior written notice.

If Convex does not cure a material breach within 30 days of receiving written breach notice, the customer, if the non-breaching party, may terminate the agreement or affected Order Forms by written notice. If the customer terminates the agreement under this provision, Convex refunds the pro-rata amount of prepaid, unused fees for the terminated Order Forms' remaining term as of the effective termination date. This is a contractual exception, not a general trial refund window.

Convex alternatives

Supabase offers per-project Postgres, authentication, storage with Row Level Security, realtime services, and Edge Functions. Its Postgres foundation makes it a relevant alternative for SQL-oriented teams.

Firebase Cloud Firestore provides a NoSQL mobile/web/server database with realtime listeners and offline support for mobile and web. That documented offline behavior addresses disconnected access; it does not establish equivalent behavior in Convex.

Limitations

Transaction and execution boundaries

• External side effects: Actions are not automatically retried on errors: an external request may already have taken effect. Callers handle retries; unfinished promises may or may not complete after return.
• Separate transactions: Actions access the database indirectly via queries and mutations. Separate runQuery/runMutation calls need not share a consistent state; one combined query/mutation has a consistent transaction boundary.
• Document size: A document is limited to 1 MiB, with nesting depth limited to 16.
• Transaction size: A query or mutation can read at most 16 MiB and write at most 16 MiB. Data scanned and then excluded by a filter still counts toward reads.

The actions guide states that actions time out after 10 minutes. The limits reference instead lists 30 minutes for Convex-runtime actions and 10 minutes for Node-runtime actions, plus one second of query/mutation user code excluding database operations. These public descriptions are inconsistent.

Search and generation boundaries

Vector search is action-only approximate nearest-neighbor search using cosine similarity. Result membership is not reactive, although a separate query can keep each result's data reactive; records can change or disappear before retrieval. Tables allow four vector indexes with dimensions from 2 to 4,096.

Image/video generation is alpha with changeable request/response formats. Images use text prompts, without input-image or mask editing. Synchronous video fails after about five minutes and permits 64 MiB downloads; cancel/timeout/lost responses may still cost. Async operations expire after seven days; providers may retain videos for less time.

Self-hosting license

The backend's FSL Apache 2.0 fair-source license excludes creating a product designed to compete with hosted Convex Cloud. Code converts to Apache 2.0 two years after creation; current releases are not unrestricted Apache 2.0 from day one.

Designated Beta Features are evaluation-only, outside the contractual Platform, and carry no service commitments, support, warranty, or liability for use; Convex may discontinue them.

Data rights and model handling

Customers/licensors retain Customer Data rights and grant Convex a limited processing license for platform delivery, maintenance/support/improvement, customer-initiated integrations, and legal obligations. This is a customer-to-platform license, not a transfer of data ownership.

Convex owns Aggregated Data and may use it for legitimate business purposes, including analysis, product development, and benchmarking. It must not identify customers/users, contain their information in identifiable form, or reasonably permit reidentification.

The AI terms prohibit training/improving Convex or third-party foundation models with Customer Data without prior written customer consent. Similar inputs may produce similar outputs, with no warranty of accuracy, completeness, error-free results, or fitness for purpose.

A ZDR marker records a zero-retention catalog match at the last update. Routing must be enabled per request and fails without an eligible endpoint; the marker alone is no guarantee. Image/video/audio routes do not accept this option.

Privacy, deletion, and processing locations

Convex's privacy policy excludes application Customer Data, which is governed by customer agreements and application policies. For policy-covered information, it may collect IP/device identifiers, approximate location and activity data, using cookies and analytics tools.

Policy-covered information may be processed/stored worldwide, including in the US, and retained during service use or for collection purposes, delivery, disputes, audits, legal defenses, business purposes, enforcement, and legal obligations; no universal fixed period is stated.

Under the DPA, Convex must delete Customer Personal Data at agreement expiry or termination. Backup/archive copies are deleted per its retention schedule; legally required retained copies must be isolated and protected from further processing except as required by applicable law. The subprocessor list ties infrastructure/database services to the selected region, while listing US support, analytics, billing and AI routing, plus German community management.

Convex states AES-256 at-rest encryption for customer code, databases, files and search indexes, encrypted transit, restricted production access, and SOC 2 Type 2 compliance. Convex says businesses subject to HIPAA may process PHI on the platform provided they sign its Business Associate Agreement; these vendor statements do not establish every customer's application compliance.

FAQ

Q1. Does every Convex function have realtime and transaction guarantees?

No. Queries are cached/subscribable and mutations are transactional; external actions do not inherit those same guarantees.

Q2. Is AI Gateway included in the Free plan?

No. It requires a paid team and separately metered model use. Self-hosted deployments use their own provider credentials.

Q3. Is selecting a European region the same as keeping every kind of data there?

No. Deployment infrastructure, policy-covered personal information, subprocessors, and model-routing retention each have different scopes.