Skip to content

itscatalyst/merakicore

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

38 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Meraki Core

Meraki Core is a local-first learning engine for AI agents.

It turns an explicit correction into immutable evidence, proposes a narrow rule, waits for approval, and returns that rule only when a later task matches its subject, project, task type, and mode.

This repository is early. The local engine, HTTP adapter, and MCP adapter work. It is not a hosted service, a model-training system, or an autonomous memory that silently rewrites itself.

One concrete example

An agent writes a vague email subject. You correct it:

Use a concise, concrete subject line.

Meraki records that correction without changing future behavior. It proposes a scoped candidate. After approval, a matching email task receives the rule. A creative-writing task does not. The evidence, rule version, pack hash, and decision remain inspectable.

Run that complete flow:

git clone https://github.com/itscatalyst/merakicore.git
cd merakicore
corepack enable
corepack pnpm install --frozen-lockfile
corepack pnpm build
corepack pnpm example:quickstart

The example proves:

  • a candidate is inactive before approval;
  • approved guidance changes a related run;
  • unrelated work stays at baseline;
  • local state survives restart.

Run the adapters

Copy .env.example values into your shell, then generate a local bearer token:

corepack pnpm dev:token

Set the printed value as MERAKI_MCP_TOKEN for MCP clients or send it as Authorization: Bearer <token> to the API.

corepack pnpm dev:api
corepack pnpm dev:mcp

Both adapters use .meraki/runtime.json by default. Do not run multiple writers against the same local file.

Repository map

apps/
  api/             authenticated HTTP adapter
  mcp/             bounded MCP adapter
packages/
  core/            evidence, profile, guidance, learning, evaluation, runtime
  contracts/       generated wire types
  auth/            signed JWT authority
  storage-local/   atomic local JSON persistence
schemas/           canonical wire contracts
examples/          executable correction-to-guidance walkthrough

The dependency direction is deliberate: applications depend on packages. MCP does not import API internals.

Verify

corepack pnpm verify

This builds and typechecks the clean workspace, checks formatting and lint, runs unit and integration tests, verifies schema/OpenAPI drift, and executes the quickstart.

Honest boundary

Meraki Core does not currently provide:

  • hosted or multi-process persistence;
  • PostgreSQL, teams, billing, or public signup;
  • a production identity provider;
  • a supported Studio UI;
  • automatic extraction from everything a user does;
  • proof that the approach works at useful scale.

Those return only after the local correction-to-guidance loop creates repeatable value.

Read Architecture, HTTP API, MCP, and Limitations for the technical boundary.

License

Apache-2.0. See LICENSE.

About

Continuous learning for AI: evidence-backed corrections, scoped guidance, evaluation, and rollback.

Topics

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages