Wren AI pricing explained — what a credit costs, and what the annual toggle hides

Wren AI pricing at a glance (read 2 August 2026)

There are four commercial plans plus a free open-source edition. The first three run on Wren AI Cloud and are metered in credits; the fourth is self-hosted and licensed by concurrent sessions instead.

• Free — $0. "20 monthly free credits", plus "80 credits for the first 14 days". Capped at 2 projects and 2 members, 10 tables per project, 2 dashboards per project and 3 AI-powered spreadsheets per project. Free credits expire: "Free credits are available for testing new features and will expire at the end of the month; they do not roll over."
• Essential Cloud — $179 per month billed annually, or $259 per month billed monthly. Includes 1,100 monthly credits (shown as "13,200 Annual credits included" on the annual view). Unlimited projects and members, the Embedded AI API, API token management, dbt integration and Basic role-based access control.
• Enterprise Cloud — $559 per month billed annually, or $799 per month billed monthly, with a "Most Popular" badge. Includes 2,000 monthly credits (24,000 annual). Adds row- and column-level data control, MCP integration, Advanced RBAC, audit logs and groups.
• Enterprise Plus — no published price; the CTA is "Talk to sales" and the card says "Annual Contract Only" and "Starts at 10 concurrent sessions". This is the self-hosted tier, "incl. on-prem & air-gapped", and the only tier where "Bring Your Own LLM" is ticked.
• Open Source — free. The card reads "Open Source · CLI & Engine" and "Built for developers; no UI." Wren's open-core page calls it "Open source · Individuals · Free forever" and "For individual developers, used through the CLI and MCP server (no UI)."
• Every plan carries SOC 2 Type II in the comparison table, including Free. The free 20 credits a month are not exclusive to the free plan: "+ 20 Monthly Credits For All Plans … This offer is available with all plans!"

Why you will see $179 quoted everywhere and be charged $259

Wren's pricing page has an Annually / Monthly toggle and it opens on Annually. Every price card therefore reads "$179" or "$559" followed by "/month, billed annually". Wren labels that accurately — the words "billed annually" are right there under the number.

What is unusual is where the monthly number lives. We downloaded the page on 2 August 2026 and searched the whole document: "$259" appears exactly once and "$799" exactly once, and both occurrences are inside the page's own JSON data payload, as "price":"$259","annualPrice":"$179" and "price":"$799","annualPrice":"$559". Neither string appears anywhere in the rendered markup. The rendered markup contains $179 and $559 only.

The practical consequence is worth knowing before a budget conversation. Anything that reads the page without running its JavaScript and clicking the toggle — a scraper, a price-comparison listing, an AI assistant summarising the page — sees $179 and $559. Those are the annual rates. Monthly, Essential Cloud is $259 and Enterprise Cloud is $799, which is $80 and $240 a month more.

The annual commitment is a genuine saving, and a large one: $179 against $259 is 30.9% off, and $559 against $799 is 30.0% off. It is just a twelve-month commitment, and Wren's own FAQ is explicit that you cannot reverse it mid-term — "Downgrading (Annual → Monthly): You must wait until the end of your current annual billing period before switching to monthly billing."

A credit is worth exactly ten cents — and Wren tells you twice without saying so

This is the most useful thing on the page and it is split across two tables that are never shown at the same time. Wren publishes its web usage in credits per million tokens, and its Embedded API usage in US dollars per million tokens. The two sets of numbers are the same prices in different units.

In Classic mode, asking a question costs 20 credits per million input tokens and 80 credits per million output tokens. The Classic API price list, for the same work through /api/v1/ask, is US$2 per million input tokens and US$8 per million output tokens. That gives a rate of exactly $0.10 per credit — which is also the overage price Wren prints on the plan cards, "$0.1 per additional credit".

It holds on every row, not just that one. Agentic mode is priced per model and per context length, and each of its five rows lines up with the dollar figure Wren publishes for the same model in its API table — 24.5 credits against US$2.45, 35 against US$3.5, 70 against US$7, 140 against US$14, and the output and cached-input rows likewise. Seventeen pairs in total, one exchange rate, no exceptions.

So you can convert any Wren credit figure into money: multiply by ten cents. The free plan's 20 monthly credits are $2 of metered usage. The 14-day welcome grant of 80 credits is $8. Essential Cloud's 1,100 monthly credits are $110 of usage inside a $259 monthly bill.

What you cannot reliably do is convert credits into a number of questions, and Wren's own material is divided on whether that is even the right unit. The pricing page meters tokens: its usage table is headed "Credits per 1M tokens". The billing documentation describes something else — "Your question quota is based on the total number of questions asked by all users in your organization's projects", with "A question only counts towards the quota if a result is selected from the query results." A token meter and a question quota are not the same instrument, and the same billing page also has a "Token-based consumption" section, so both descriptions are live. What is certain is that a question's cost varies with its size: a small schema and a short SQL statement is a fraction of a credit, the same question against a wide warehouse with a long conversation behind it is many times more. Wren publishes no typical tokens-per-question figure, so any page telling you the free tier is worth "N questions" has supplied that number itself.

Two smaller mechanics worth knowing, both from Wren's own pages: failed work is not charged — "Credits are only deducted when an operation is successfully completed… If an action fails or returns an error, no credits will be charged" — and free credits are spent before the ones you paid for: "Free credits are deducted first before subscription-based credits." In agentic mode there is also a quality dial that moves the meter: "Higher effort means the agent thinks longer before answering, which improves quality on hard questions at the cost of latency and credits."

There are two credit pools, and the pricing page only tells you about one

This is the largest gap between Wren's pricing page and Wren's documentation, and it is the one most likely to change a purchase decision.

The pricing page lists "Embedded AI API" and "API Management" as ticked features of Essential Cloud and above, in the same visual language as every other feature. Its FAQ describes API usage as drawing on "your credits". Read that way, the API is a capability your subscription unlocks and your monthly credit pool pays for.

The billing documentation says otherwise, in its own opening summary: "Wren AI uses two credit pools: Web credits — included with your subscription and consumed when people ask questions in the Wren AI app. API credits — purchased separately and consumed by requests to the Wren AI API." And again, explicitly: "API credits are separate from web credits and are required for making requests to the Wren AI API."

So on the cloud plans, the API is a second bill. The docs describe it as a prepaid dollar balance with its own controls: "Auto Recharge: You can enable or disable auto recharge. If auto recharge is off and your balance reaches $0, API requests will not be processed." The configurable thresholds have floors — for the balance trigger, "The minimum allowed is $5", and for the top-up, "The minimum allowed is $15." None of that appears on the pricing page.

The consequence reaches further than a developer integration, because one of the plan features bills to the API pool. Wren's Slack documentation states it plainly: "Each question uses API credits for an 'Ask Question' (/api/v1/ask) request. Each chart generation uses API credits for a 'Generate Chart' (/api/v1/generate_vega_chart) request." Slack Integration is presented on the pricing page as a plan feature — "3 Channels" on Essential Cloud, ticked above that — with no indication that questions asked there draw on a balance you have to buy separately from the subscription. If your plan is to roll Wren out through Slack, budget for both pools.

The self-hosted tier is the exception, and Wren is clear about it: "With concurrent session licensing there is no separate per-call API charge to calculate", and API requests "count toward your concurrent session license, exactly like usage in the UI, Teams, or MCP." One licence, one meter.

One thing we could not resolve: whether the two pools are interchangeable in either direction. Neither the pricing page nor the billing documentation says, so we are not going to guess.

The credits in your plan cost more than the credits you buy when you run out

Divide each plan's price by the credits it includes and you get a per-credit figure. On every plan and every billing period, that figure is higher than the $0.10 you pay for extra credits.

This is not a trick and it is not a criticism — it just means the subscription is not mostly buying credit volume. What the money buys is the plan's features and its limits: unlimited projects and members instead of two and two, CSV download, the Embedded API and API tokens, dbt, RBAC, and on Enterprise Cloud the row- and column-level security, MCP and audit logs. The credits are bundled in, priced as part of the plan rather than at cost.

It is still worth knowing which way the gradient runs, because it inverts the usual intuition. Most metered products punish overage. Here, if your usage is spiky, the marginal credit is the cheapest credit Wren sells, and the plan you should pick is the cheapest one whose features and limits you actually need — not the one whose credit bucket looks big enough. Wren's own FAQ confirms extras are simply purchasable: "If you exceed the usage limits for your credits, you need to buy additional credits in the billing section."

One caveat we could not resolve from the published material: Wren does not state whether extra credits can be bought on the Free plan. The $0.1 line appears on the Essential and Enterprise Cloud cards and not on the Free card, so we have written "Not offered" for Free in the table above rather than asserting it either way.

Which mode you use matters more than which plan you buy

Wren has two answer modes, Classic and Agentic, and both are ticked for every plan including Free. Their token rates are not close.

Classic mode uses "the default model" and costs 20 credits per million input tokens and 80 per million output. Agentic mode is priced per model and then again per context length, with a break at 272K tokens. At the cheap end, gpt-5.2 is 24.5 in and 196 out — 1.2x the Classic input rate and 2.45x the output. At the expensive end, gpt-5.5 above 272K of context is 140 in and 630 out, which is 7x and 7.9x Classic.

The context break matters as much as the model choice, because crossing it doubles the rate. On gpt-5.4, short context is 35 credits in and 210 out; long context is 70 and 315. On gpt-5.5, short is 70 and 420; long is 140 and 630. In other words gpt-5.5 on short context costs exactly the same input rate as gpt-5.4 on long context. Which side of 272K your conversation and schema land on is a cost decision you make without being asked.

Put that against the free allowance and the spread becomes concrete. Twenty credits buys a million input tokens in Classic mode, about 816,000 on gpt-5.2, about 571,000 on gpt-5.4 short, about 286,000 on gpt-5.4 long or gpt-5.5 short, and about 143,000 on gpt-5.5 long. Same plan, same allowance, a seven-fold difference in how far it goes, decided by two toggles rather than by a purchase.

Cached input is the other lever and it is a big one: Wren prices cached tokens at exactly one tenth of fresh input on every agentic row (2.45 against 24.5, 3.5 against 35, 7 against 70, 14 against 140). Repeat and follow-up work over the same context is therefore an order of magnitude cheaper than first-touch work, which is the opposite of how a per-question meter would behave.

If you are evaluating Wren on cost, model the mode, the model, the context length and the caching — in that order — and only then the plan. The plan sets your limits; those four settings set your burn rate.

Where Wren's own pages disagree with each other

This is the section an aggregator listing cannot write, because it requires reading Wren's pricing page, its open-core page, its documentation and its repository side by side on one day. None of these are accusations of bad faith — every fast-moving product accumulates drift between its marketing page, its docs and its code. But each one is a place where you could plan against the wrong fact.

• Roll-over is described three different ways. The Essential and Enterprise Cloud cards both read "Credits roll over up to 2x ($0.1 per additional credit)", and the comparison table's pricing-model row repeats it unqualified: "Cloud plans are usage-based: a monthly pool of credits (roll over up to 2x, $0.1 per extra credit)." The FAQ lower down the same page narrows it to one billing period: "the roll-over mechanism applies only to monthly plans; unused credits in an annual plan do not roll over across billing cycles", and "Annual Plan: Credits do not roll over and will be fully reset at the start of each new billing cycle." The billing documentation then removes it entirely: "All web credits are granted upfront. Any unused web credits expire when the current subscription period ends." We searched that page for "roll over" and "rollover" and found neither phrase. Since the pricing page opens on the annual view, the cards you see first advertise a benefit that its own FAQ excludes from the plan being displayed and its own docs do not mention at all.
• "All in Enterprise Cloud, PLUS" is not literally a superset. The Enterprise Plus card opens with that line, but the comparison table on the same page ticks CSV Upload for Enterprise Cloud and marks it "Coming Soon" for Enterprise Plus. If uploading a spreadsheet matters to you, the most expensive tier is currently the one that cannot do it.
• The Enterprise Cloud card and the comparison table disagree about LDAP and OIDC. The card lists "Custom LDAP/ AD & OIDC" among what that tier includes; the table marks both LDAP/AD and OpenID Connect as "Add-on" for Enterprise Cloud, and ticks them outright only for Enterprise Plus. An add-on is normally a separate purchase.
• The connector count comes in five sizes. The documentation home page says "17+ data sources"; the pricing page footer and the open-core page say "20+ connectors" and "20+ data sources"; the repository README says "22+ data sources" in three separate places; the cloud connector overview lists fourteen entries; and the open-source connect guide enumerates thirteen — DuckDB, PostgreSQL, MySQL, BigQuery, Snowflake, ClickHouse, Trino/Presto, SQL Server, Databricks, Redshift, Oracle, Athena and Spark. The likeliest reading is that the enumerated lists count connectors while the headline counts count something broader, but Wren does not say so, and we are not going to state it as fact. If a specific engine is why you are buying, check that engine's own connector page rather than any of the counts.
• The documentation says users and projects are unlimited; the pricing page caps the free plan at two of each. Wren's billing page states flatly: "The number of users and projects per organization is unlimited. Feel free to invite as many teammates as you need." The pricing table gives Free 2 projects and 2 members, and two other documentation pages agree with the pricing page rather than with the billing page — "Free plan: Up to 2 projects" and "Free Plan: Limited to 2 members (including Admin)." The billing page's sentence is true of the paid plans and wrong about Free. Separately, the pricing page's 10-tables-per-project figure is a Free-column value, while the modelling docs give a general ceiling with no plan qualifier: "Each project in Wren AI supports importing up to 200 models (tables) by default. This is a soft limitation — you can increase this limit by contacting us."
• Three plan names appear in the documentation that are not on the pricing page. A "Starter" plan is referenced in the agentic-mode docs, which describe a one-time grant of "Starter: 100, Essential Cloud and Enterprise Cloud: 500" credits. A self-hosted "Business plan" gates or grandfathers features on more than ten documentation pages, always in the same form — "Existing Business plan customers keep access" — and only one page explains it: "The self-hosted Business plan is no longer offered to new customers — new self-hosted deployments use Enterprise Plus." And a grouping called "Wren AI Paid plans" gates several features in the open-source docs. None of the three is a tier on the pricing page. If you are reading a feature page to find out whether your plan includes something, the plan it names may not be one you can buy.
• The saving on annual billing is real and never stated. We found no discount badge, no "save 20%", no "two months free" and no "billed monthly" comparison anywhere in the pricing page. The 30.9% and 30.0% figures in this guide are our own arithmetic on Wren's two published prices. That is an unusual thing for a vendor to leave on the table, and it is the same root cause as the missing monthly price: the annual view is the only one the served page contains. Requesting the page with ?tab=monthly returns byte-identical HTML, so the monthly view exists only in the browser.
• The licence is described three ways. GitHub's API reports the repository licence as "Other" with SPDX "NOASSERTION". The LICENSE file itself opens "WrenAI is multi-licensed. The license that applies depends on the path within this repository" and maps core/, sdk/, skills/ and examples/ to Apache 2.0 and docs/ to CC BY 4.0, while shipping an AGPL-3.0 text it describes as provided "pre-emptively" for modules that do not exist yet. The README simply says "Apache 2.0" with a matching badge. If your legal review needs one answer, the LICENSE file is the one with the detail — and note separately that "The names \"Wren\", \"WrenAI\", and the project's logos are trademarks of Canner, Inc. and are not licensed under any of the licenses referenced above."
• MCP is described three ways at once. The open-core page puts the "MCP server" inside the open-source layer and calls that layer "Free forever", and the open-source docs document the command that runs it: "wren serve mcp exposes a Wren project's query, schema, and business-knowledge". But the open-source-versus-commercial comparison table in the same documentation marks "MCP server and hosted REST API" with a cross for open source. And in the cloud pricing table, MCP Integration is blank for Free and Essential Cloud and ticked only from Enterprise Cloud upward. Our reading is that the self-hosted MCP server is open and the hosted one is not, which would make all three statements compatible — but the comparison table does not draw that distinction, and if MCP is why you are choosing Wren, test the command before you choose a plan.
• The open-source card says "no UI"; the documentation ticks dashboards for open source. The pricing card is "Built for developers; no UI." The open-source-versus-commercial table marks "Web UI for non-technical users" with a cross for open source, which agrees — and on the very next line marks "GenBI dashboards" with a tick for open source, which does not. We could not tell from the published material what a dashboard without a UI looks like in the open-source edition.
• The phrase "Bring Your Own LLM" does not appear in the repository at all. It is on the pricing page twice — as a bullet on the Open Source card and as the one row ticked for Enterprise Plus. We searched all 144 markdown files on the default branch for "Bring Your Own LLM" and for "BYO" and found zero matches, and the only API key documented in the current operational reference is OPENAI_API_KEY, "Required for memory embeddings". What Wren does publish is elsewhere and in two different places: the commercial bring-your-own-LLM docs list four templates — OpenAI, Azure OpenAI, Google Vertex AI and AWS Bedrock — and the sunset Classic documentation takes a different approach entirely, "We are now using LiteLLM to support LLMs, so you can use any LLM supported by LiteLLM." Those are two products with two different answers, and the current open-source install guide gives neither. If you need a specific provider on the current open-source edition, ask before you build on it.

What the Free plan actually gives you, and where it stops

Wren's free tier is unusually good at one thing and unusually strict at another, and the difference decides whether it is worth your afternoon.

• It reaches your real data, and that is rare. "Database & Data Warehouse Connectors" is ticked in the Free column. You are not exporting a spreadsheet to a chat window; you point it at the warehouse where your data already lives. Among free tiers in this category that is the meaningful dividing line.
• The credit allowance is the binding limit, and it is small in absolute terms: 20 credits a month, which at Wren's own $0.1 overage rate is $2 of metered usage, plus an $8-equivalent grant for the first 14 days. In Classic mode that is a million input tokens a month. In Agentic mode on the largest model it is about 143,000. Unused free credits are lost: they "will expire at the end of the month; they do not roll over."
• The structural caps bite next: 2 projects, 2 members, 10 tables per project in the semantic modelling UI, 2 dashboards per project, 3 AI-powered spreadsheets per project. The tables-per-project cap is the one to check first — a 10-table ceiling on the semantic layer is a real constraint on anything beyond a focused evaluation.
• You can get data in but not out. CSV Upload is ticked for Free; CSV Download is blank, and so is "Export as SVG, PNG". You can load and query, but the free tier does not hand you a file back.
• No API, no dbt, no MCP, no Slack. Embedded AI API, API Management and dbt Integration all start at Essential Cloud; MCP Integration starts at Enterprise Cloud, and Slack at Essential. If your plan is to build something on top of Wren, the free tier is not the place to prototype it — the open-source CLI is. And note from the section above that reaching the API on a paid cloud plan means buying a second, separate credit balance, not spending the one in your plan.
• Two things Free does get that you might not expect: both answer modes, Classic and Agentic, are ticked for Free, and so is SOC 2 Type II. There is no card requirement anywhere on the page, and no trial expiry on Free — it is a standing free plan rather than a countdown.

The open-source edition is free, but check which Wren you are downloading

Wren describes itself as open core, and the boundary is published rather than implied: "Open source: the context engine for individual developers, used through the CLI and MCP server (no UI), with MDL modeling, a core text-to-SQL pipeline, and 20+ connectors. Commercial: the advanced AI harnesses (GenBI Apps, agentic mode, AI-assisted context preparation), the UI and dashboards, embedded APIs, access control with user and group management, row- and column-level security, advanced security, and support." That is a clearer statement of an open-core boundary than most vendors publish, and it is worth reading before you assume the free download is the product in the screenshots.

The critical thing to check is which version you are looking at, because the project restructured in 2026 and most third-party write-ups still describe the old one. The engine merged into the main repository under core/ on 7 May 2026. The Docker-based chat-first BI application — the one with a web UI that older tutorials walk you through — was moved to the legacy/v1 branch and renamed Wren GenBI Classic, and Wren's own documentation calls it sunset: "Its code lives on the legacy/v1 branch and no security fixes will be issued."

That matters more than a naming change. We read the branch on 2 August 2026: legacy/v1's last commit is dated 4 May 2026, and its message is "fix(wren-ai-service): upgrade litellm to 1.83.7 (SQL injection CVE)". The final change to that branch was a security fix, and Wren has said there will not be another. If you follow a 2025 tutorial and docker-compose up the old stack, you are running code with a published no-more-fixes policy.

What you install today is different in kind. There is no docker-compose.yaml on the default branch at all — we checked six likely paths and the root listing, and the Docker and deployment directories exist only on legacy/v1. The current install is pip install wrenai, with per-source extras such as pip install "wrenai[postgres,memory]", plus an agent-skill installer, npx skills add Canner/WrenAI, that targets coding agents like Claude Code, Cursor and Codex. You drive it with commands like wren query and wren ask. It is a developer tool for building your own agents, exactly as the pricing card says: "Built for developers; no UI."

The trap is that both installation guides are still live and neither one warns you off the other. One page tells you to run npx skills add Canner/WrenAI; another tells you to "Install Docker Desktop version 4.17 or later" and pins launcher binaries to release 0.29.1. The Docker page sits under a sidebar heading marked "Wren GenBI Classic · Sunset", but the page body itself carries no deprecation notice — the sunset status is stated on a different page. If you arrive from a search result rather than from the sidebar, there is nothing on the page to tell you that what you are about to deploy has a published no-more-security-fixes policy. That is the single most expensive mistake available here, and it takes one click on the sidebar heading to avoid.

The project itself is healthy, which is the other half of the question and worth stating plainly because it contrasts with other open-source options in this category. Reading GitHub's API on 2 August 2026: not archived, 16,774 stars, 1,895 forks, 359 open issues, and pushed the same day we looked. Releases are per-package rather than one product version — wren-v0.13.2, wren-core-py-v0.7.3 and so on in the days before we read it. Wren's own answer to the abandonment question is "Judge the cadence, not the promise", and the cadence supports it. Note also that the predecessor repository Canner/wren-engine is now archived, as the README states, so a link to it in an older article is a dead end rather than a second project.

How Intellrise prices the same job

• Free — $0, permanently, no card. One database source plus Google Sheets, 3 dashboards with unlimited charts, unlimited reports and every export format (CSV, Excel, PDF, PPTX, DOCX), AI report editing, custom instructions and dbt schema editing. One seat.
• Pro — $29/month, or about $24/month billed annually. Multiple data sources plus CSV and Excel upload, unlimited dashboards with full customisation, AI Semantic Schema Learning, per-key AI configuration, read-only share links and email delivery. Every new account starts on a 14-day Pro trial with no card, and the Free plan is the landing state afterwards rather than an expiring demo.
• Bring-your-own-key on every tier, including Free — Gemini, OpenAI, Anthropic, DeepSeek, MiniMax or any OpenAI-compatible endpoint. There is no credit meter, no per-question markup and no token accounting on our side at all: you pay your provider directly and we never see the bill.
• The comparison at one seat, honestly stated: Intellrise Pro is $29/month plus whatever your own model calls cost. Wren AI Essential Cloud is $259/month monthly or $179/month annual, with $110 of credits included per month. Those are different products at a 6-9x price gap, and the gap is not arbitrary — the sections above and below say what the extra money buys.
• Where the shapes genuinely differ: our meter is structural, not per-token. Hitting our Free limits blocks the next data source or the next dashboard, never the next question. Wren's cloud meter is per million tokens, so the cost of asking one more follow-up is never zero. Neither shape is better in the abstract; they are better for different usage patterns. If your team asks a few hundred questions a month over a stable set of sources, a flat plan plus your own key is cheaper and more predictable. If your usage is occasional and bursty, a credit pool you top up at ten cents is the more efficient shape.

Where Wren AI is the better buy — and for a lot of teams it is

A comparison that concedes nothing is not worth reading, so here is the concession, and it is substantial.

If your data cannot leave your network, this is not a close call. Wren's Enterprise Plus runs self-hosted, on-premise, and fully air-gapped, and the open-source engine runs entirely on your own machine. Intellrise is cloud-only. We have no self-hosted edition, no on-premise deployment and no air-gapped option, and no plan to pretend otherwise: if you are in a regulated environment with a hard data boundary, we are not a candidate and Wren is.

If you want to read the code, likewise. Wren's engine is on GitHub and every commercial plan runs that same engine. Intellrise is not open source — you cannot read our query pipeline, and "trust us" is the honest description of what we offer on that axis. Wren's argument here is a good one and we are not going to argue with it.

If you need governance, Wren is ahead of us on every line. Row- and column-level data control, groups, audit logs, Advanced RBAC, SSO with SCIM 2.0, LDAP/AD and OIDC, SOC 2 Type II on every tier including Free. We have none of those. Our Team plan — shared workspace, roles and seat management — is marked "soon" on our own pricing page with a notify-me link, not a buy button, and it would be dishonest to count it. If more than one person needs the same sources today, that alone decides it.

If you want your definitions as code, Wren is the design that does that. Its Modeling Definition Language keeps models, relationships, metrics and access rules in files in your own git repository, reviewable in a pull request, and Wren's lock-in answer follows from it: "Your MDL files already live in your git repo … the open-source engine keeps running without a license." Our semantic layer takes the opposite approach — the AI annotates your schema and learns definitions from your questions, and you inspect and correct rather than author. That is lower effort to start and it is not reviewable in a diff. Which one you want is a real architectural preference, not a feature gap.

And if you are building a product on top, Wren has the surface for it: an embedded AI API with published per-token pricing, API token management, an MCP server, framework SDKs for LangChain and Pydantic AI, and a WebAssembly build. We have no public API.

Where the flat-plus-your-own-key shape wins is narrower and worth stating precisely: when the people who need answers do not write SQL and will not maintain a modelling language, when the deliverable is a dashboard or an exportable report rather than an engine call, when nobody is available to operate a self-hosted stack, and when you want the marginal question to cost nothing rather than tick a meter. Outside that, and especially inside a compliance boundary, buy Wren.

Check it on your own data before you pay for either

Wren's free cloud plan connects to a real database and needs no card, and the open-source CLI is a pip install away, so you can evaluate it properly without spending anything. Do it with your real warehouse rather than a sample, and watch two numbers: how far 20 credits actually goes on your own question sizes, and whether ten tables per project is enough for the semantic model you would need. Those are the two limits most likely to surprise you, and both are cheap to test.

On our side there is a public demo at https://intellrise.com/demo/ that needs no signup, no card and no AI key — a real dataset, the SQL that was generated, and the output it produced — so you can judge the substance before connecting anything. If it looks right, every new account starts on a 14-day Pro trial with no card, and the Free plan keeps working afterwards: one database source plus Google Sheets, three dashboards, and unlimited reports and exports on your own AI key.

Two honest caveats on ours, both of which cost us evaluations. Bring-your-own-key means a setup step before anything runs: you have to create a provider key first, and Wren's free cloud tier does not ask that of you. And we connect ten sources — PostgreSQL, MySQL, SQL Server, BigQuery, Snowflake, Databricks, Redshift, Google Sheets, CSV and Excel — against Wren's 20-plus. If your data lives in DuckDB, ClickHouse, Trino, Oracle, Athena or Spark, we cannot read it at all and Wren can.