Open source · v0.1.0-alpha.0 · Apache-2.0
Write down the rules of your software once: who is allowed to do what, and what must always be true. Averri turns them into facts your AI coding agent checks instead of guessing, and tests that prove they hold.
npx -y @averri/cli mcp
Works with Claude Code, Cursor and any MCP-compatible agent.
Why Averri exists
AI assistants now write a large share of the world's software. They are fast and talented. But every task starts like a first day at work: they don't know the unwritten rules.
Nobody wrote these down. They live in people's heads. So the AI does the job well and sometimes quietly breaks a rule it never knew existed. A smarter AI is still a new hire with no handbook.
Averri is the handbook, and the building inspector. You write the rules once. The AI reads them before it changes anything, and automatic tests check them every time.
Coding agents are good at code that compiles. They're weak at invariants that live outside the code: authorization, idempotency, pre- and postconditions, event guarantees.
Averri adds a typed declaration layer to an existing Prisma + Hono/Express + Zod backend. A static compiler turns declarations into a semantic graph with provenance (extracted, declared, verified), emits structured diagnostics, and generates Vitest suites. Agents query it over MCP: inspect before editing, impact before changing, check before finishing.
No new syntax, no runtime to host, no application code generated. Remove the imports and your app is exactly what it was.
Try a rule
Pick who is trying. This is exactly the check Averri writes a test for, automatically, from one line of rule.
Choose someone
Averri generates one test per row of this table, so nobody has to remember to.
How it works
Next to the routes you already have. Who may call it, what must be true before and after, which event must fire.
export const InviteMember = action("Team.inviteMember", {
route: { method: "POST", path: "/teams/:teamId/invitations" },
policy: OwnerOrAdmin,
requires: { notAlreadyMember: db.none("membership", { … }) },
ensures: { eventEmitted: e.emitted(MemberInvited, 1) },
});It reads your Prisma schema, routes, Zod schemas and rules, without ever running your code, and checks them against each other.
Your AI agent asks Averri before it edits, and generated tests check every rule on every change.
overviewWhere am I?findWhere is X?inspectWhat is true about this?impactWhat breaks if this changes?checkIs my change correct?explainWhy did this fail?read_sourceShow me the code.suggestDraft rules for me.Proof, not promises
On our 38-route example app, built by an AI agent, the "view billing" page was quietly writing to the database. The app ran fine. The normal tests passed. Averri's compiler compared what the page claimed to do with what its code actually did, and flagged it.
An AI wrote the bug. A tool that gives AI the rules caught it.
For developers
npm install --save-dev @averri/cli @averri/core @averri/contracts
npx averri init # detects Prisma + Hono/Express + Zod, writes averri.config.ts
npx averri build # builds the graph and reports diagnostics
npx averri testgen # writes generated tests to tests/averri/
npx averri check # typecheck + affected tests, structured outputclaude mcp add averri -- npx -y @averri/cli mcpRun it inside your project. averri init also adds a short block to CLAUDE.md telling the agent when to use each tool.
{ "mcpServers": { "averri": { "command": "npx", "args": ["-y", "@averri/cli", "mcp"] } } }Save as .cursor/mcp.json in your project.
npx -y @averri/cli mcp # stdio
npx -y @averri/cli mcp --http 3333 # streamable HTTP
npx averri overview --json # same answers from the command lineAverri never executes your code at build time and sends no telemetry. Ever.
The story
What if a programming language were designed for AI agents first? Fewer wasted tokens, self-explaining errors, software that carries its own map.
A research sweep, three competing strategies and six critics. It killed most of the idea: a brand-new language is the one language AI has never seen, so it makes more mistakes.
AI already writes code that runs. What it can't know is the rules nobody wrote down. That gap doesn't close as models get smarter, because the rules were never in the code.
About 70% confidence each, as a business. Crowded market, slow revenue. They were probably right about the business.
No company, no paid tiers. A free tool, a public benchmark plan and written kill criteria: if agents don't measurably do better work with Averri, that result gets published too.
Eight packages on npm, verified from the public registry. The name took five tries: AgentLang, Warrant, Surety, Averr, Averri. From aver: to state something as true.
An open invitation
Averri was argued over, torn down, rebuilt and tested more than a thousand times before its first release. It was built with care, with AI agents, for a future where AI writes most of our software. Now it needs people who will push it harder than we can.
Questions
Averri is a free, open-source tool that lets you write down the rules of your software, such as who is allowed to do what and what must always be true, so that AI coding agents check those rules instead of guessing them, and automatically generated tests prove the rules hold.
Teams whose backend code is increasingly written by AI coding agents such as Claude Code or Cursor, and the founders, product managers, engineers and researchers who need to trust that code.
Yes. Averri is free and open source under the Apache-2.0 license. There is no paid tier, no telemetry and no hosted service.
Averri runs as an MCP (Model Context Protocol) server. Add it with one command, for example claude mcp add averri -- npx -y @averri/cli mcp, and the agent can ask Averri about rules, dependencies, the impact of a change, and whether its change passes the generated tests.
Version 0.1 supports TypeScript backends on Node.js 20 or later that use Prisma, either Hono or Express, and Zod.
No. In observe mode you add declarations next to your existing routes, and Averri checks them around your current code in tests. Nothing goes into your request path.
No. The Averri compiler reads your code statically and never executes it, and Averri sends no telemetry or usage data anywhere.
AI code reviewers infer intent from a diff. Averri works from rules a human wrote and approved, turns them into deterministic tests and graph facts, and flags any change that weakens a rule, whether a human or an AI made it.
No. Averri began as AgentLang, an idea for a new language designed for AI agents. Research showed a new language would make AI agents less accurate, so Averri uses plain TypeScript instead.
Not yet. Version 0.1.0-alpha.0 is a pre-alpha release. It is tested, with over a thousand automated tests, but its interfaces may change.
Averri was founded and curated by Abhijeeth Gadwal, an AI product manager, and built with the help of AI coding agents.