ainecto — AiNecto CLI and MCP connector

Published as ainecto, @ainecto/mcp and @ainecto/cli — the same code. Command: ainecto. The older names (ai-erd, @ai-erd/mcp, @ai-erd/cli, command ai-erd) still get the same code and still work.

In a terminal, from your repository root: npx -y ainecto@latest init — it asks which role this repository's AI sessions should have and which AiNecto project to use, and signs you in through the browser if needed.

For AI agents

Ask the user which role this repository's AI sessions should have — Design, Development, Test, or Validation. Never pick one yourself. Then, from the repository root: npx -y ainecto@latest init --role <role> (it signs the user in through the browser if needed).

• Use a long command timeout (at least 6 minutes): sign-in waits up to 5 minutes for the user.
• If init lists projects, ask the user which one and re-run with --project <uuid>. If it says this is not the repository root, re-run from the folder it names.
• No restart is needed. Right after init, keep working in the same session through the CLI: ainecto commands run from the repository read its role on every call (read .ainecto/HARNESS.md, then ainecto tools list and ainecto tools call <tool> '<json>'). Do not use MCP tools the session started with for that work — they keep the role (or no role) the session started with.
• A new session is optional and only adds the role's MCP tools (Claude Code: exit, then claude -c in this folder to keep the conversation; approve the ainecto server from .mcp.json when asked). Codex: save the profile init prints and start with codex -p <role>.
• A server added during a session is not available in that session. In our test, a server added with claude mcp add while Claude Code was running did not attach to the running session — that is why the CLI is the way to keep going. /mcp shows what the current session actually loaded.
• Inside the repository, ainecto commands (ainecto tools call …) use the repository's role automatically — do not pass --role; a different --role is refused.
• Claude Code's auto mode may block running an external package with npx. If it does, ask the user to allow the command.

To connect AiNecto without a repository role (for example before init), add the remote server at user scope under the name ainecto — a local-scope entry would hide the .mcp.json that init writes, and a different name would sit next to it as a second server with no role:

claude mcp add --scope user --transport http ainecto https://ainecto.com/mcp  # Claude Code
codex mcp add ainecto --url https://ainecto.com/mcp                           # Codex

The client signs the user in on first use. That session has no role until init runs and a new session starts.

Upgrading from ai-erd (0.4.x)

Version 0.5.0 renamed what init writes into a repository. A repository set up by 0.4.x keeps working: the CLI reads the ai-erd server entry, the .ai-erd/ folder, the <!-- ai-erd:begin --> block and AI_ERD_ROLE as its own. Running npx -y ainecto@latest init again moves them to the new names — nothing is left behind twice:

| 0.4.x | 0.5.0 | | --- | --- | | server entry "ai-erd" (npx -y @ai-erd/mcp --role …) | "ainecto" (npx -y @ainecto/mcp --role …) — the old entry is replaced | | folder .ai-erd/ (HARNESS.md, config.json, init-record.json) | .ainecto/ — moved, and the old folder removed if 0.4.x created it | | block <!-- ai-erd:begin --> … <!-- ai-erd:end --> | <!-- ainecto:begin --> … — replaced in place | | AI_ERD_ROLE | AINECTO_ROLE — both are read; when both are set, AINECTO_ROLE wins |

ainecto init --undo still puts back what was there before the 0.4.x init, with or without that second init. Not renamed: the role header and OAuth scope the server reads, and the sign-in on this machine.

• If you added AiNecto earlier as ai-erd at Claude Code user scope or in Codex, remove it (claude mcp remove ai-erd -s user, codex mcp remove ai-erd) and add it as ainecto as above: under a different name it no longer hides behind the repository's entry.
• Claude Code shows the tools as mcp__ainecto__… instead of mcp__ai-erd__… — update any allow list that names the old prefix.
• An ai-erd 0.4.x installed globally does not know the new names: inside a repository set up by 0.5.0 it finds no role (so no restriction), and running its init there adds a second entry, block and folder. Update it (npm install -g ai-erd@latest gets this same code). If that already happened, 0.5.0 stops and says how to undo it (npx -y -p @ai-erd/[email protected] ai-erd init --undo).

Set up a repository (roles)

npx -y ainecto@latest init                     # in a terminal: asks for the role and the project
npx -y ainecto@latest init --role development  # agents, scripts, CI: nothing is asked

In a terminal (stdin and stdout both a TTY, no --json, and not under an agent or CI — CLAUDECODE, CI, GEMINI_CLI, CODEX_SANDBOX, CURSOR_AGENT) init asks for what was not given:

• the role, unless --role or AINECTO_ROLE is set — by number, with no default. The role already set on the repository is the default only when there is one (if its agent configs disagree, there is no default);
• the project, unless --project is set — by number; «Create a new project» is offered to Design only.

Ctrl+C or Ctrl+D at a question cancels without writing repository files (Ctrl+D exits with 130). A sign-in you already completed stays saved, and choosing «Create a new project» creates it right away. Anywhere else (an agent's shell, CI) it asks nothing. In both cases:

• If this machine is not signed in to AiNecto yet, init opens a browser and continues once the user approves — once per machine (per server), not per role. The sign-in URL is printed first (with --json, as one JSON line on stderr: {"event":"login_url","url":…}), and sign-in gives up after 5 minutes. It fails at once when the command that opens the browser is missing, or (except on Windows, where the exit code is not reliable) exits with an error.
• Outside a terminal, with no --role/AINECTO_ROLE it uses the role already set on the repository. If there is none (or the repository's agent configs disagree) it stops and says to ask the user — it never picks a role for you. If the account has more than one project it lists them and stops; re-run with --project <uuid>. With no project yet, only a Design session can create one (--yes, optionally --project-name <name>); in any other role it asks you to have the user create the project, then re-run with the same role.
• It stops when run from a sub-folder of a git repository, and tells you the root.
• It writes one ainecto server entry carrying --role into .mcp.json (Claude Code) and .cursor/mcp.json (Cursor), and prints the Codex profile to use (codex -p <role>).
• ainecto commands in the repository use the new role at once; MCP tools pick it up in a new session (a running session's MCP tools keep the role they started with). Roles are a working guardrail, not a security boundary.
• ainecto init --dry-run shows what would change (it does not sign in or refresh a sign-in); ainecto init --undo removes what it wrote.

Commands

Prefer npx -y ainecto@latest <command>. To install once instead, use npm install -g ainecto — but install only one of ainecto, @ainecto/mcp, @ainecto/cli, ai-erd, @ai-erd/mcp and @ai-erd/cli globally: they are the same code and ship the same commands, so a second global install conflicts. (ainecto and ai-erd carry both commands, ainecto and ai-erd; @ainecto/cli carries only ainecto, @ai-erd/cli only ai-erd.)

ainecto auth login          # once per machine (and per --env); not per role
ainecto auth status
ainecto auth logout

ainecto tools list
ainecto tools call list_projects --json
ainecto tools call erd_apply_changes -f changes.json --json
cat payload.json | ainecto tools call erd_apply_changes --json

ainecto projects list
ainecto erd apply-changes -f erd-operations.json --yes

ai-erd is the same command under its older name — every ainecto … line works as ai-erd ….

The role comes from the repository. Inside a folder set up with ainecto init (the command looks from the current folder upward for the first .mcp.json/.cursor/mcp.json with an ainecto entry — or, from 0.4.x, an ai-erd entry), every command uses that folder's role without --role. A --role or AINECTO_ROLE (or AI_ERD_ROLE) that differs from it is refused — the role changes only with ainecto init --role <role>, after asking the user. If the role there cannot be read (the @ainecto/mcp bridge entry without a valid --role, configs that disagree, an ainecto and an ai-erd bridge with different roles, broken JSON that mentions ainecto or ai-erd, a link that points outside the folder), the command stops. An entry that is not this bridge (for example an HTTP url or mcp-remote entry) means no role there. Your home folder itself is never read as a repository (~/.cursor/mcp.json is Cursor's global config). Outside such a folder there is no role (no restriction), and --role narrows it. Tool names are the server's own names (list_projects, erd_apply_changes, …). Your AI client shows them with its own prefix, for example mcp__ainecto__list_projects in Claude Code (mcp__ai-erd__… for a 0.4.x setup); the CLI does not use that prefix. A tool name that starts with the server's older mcp__ainecto__ prefix is still accepted.

Every command takes --env dev for https://dev.ai-erd.com/mcp. Endpoint resolution priority:

1. --endpoint <url>
2. AINECTO_MCP_ENDPOINT
3. --env dev
4. production default, https://ainecto.com/mcp

Endpoint URLs must use https:. Plain http: is accepted only for localhost loopback targets such as 127.0.0.1. When AINECTO_TOKEN is set, the CLI only sends it to the default prod/dev endpoints unless AINECTO_ALLOW_CUSTOM_ENDPOINT_TOKEN=1 is set for an explicitly trusted custom endpoint.

The commands are ainecto and ai-erd (the same CLI), ai-erd-mcp, and mcp (which npx -y @ainecto/mcp runs). The AINECTO_* environment variables above are the only names for those settings — there is no AI_ERD_* equivalent, so do not guess one. (The one AI_ERD_* variable is AI_ERD_ROLE, the older name of AINECTO_ROLE.)

Generated Friendly Commands

Every checked-in generated catalog tool is reachable through its deterministic command path. Scalar schema fields are exposed as flags, with both kebab-case and schema-case accepted:

ainecto documents list --project-uuid <projectUuid>
ainecto documents list --projectUuid <projectUuid>

Array or object payloads use the first-party CLI JSON payload reader:

ainecto documents create -f create-documents.json
cat erd-operations.json | ainecto erd apply-changes --json

Destructive generated commands prompt in human mode unless --yes is supplied. In --json mode they fail with a structured error unless --yes is present.

Attachment file upload is available through a bespoke command that performs the upload-token, raw PUT, and attachment registration flow:

ainecto attachments upload --document-uuid <documentUuid> ./diagram.png ./notes.pdf

The generated request_upload_token and upload_attachments paths remain raw MCP argument-contract commands and do not read local files.

Connector Mode

ainecto mcp --role <design|development|test|validation> [--env dev]

The package doubles as a stdio-to-HTTP proxy for the resolved /mcp endpoint. It proxies initialize, tools/list, and tools/call, does not rewrite remote schemas, and does not interpret local file references. It uses this machine's one sign-in (ainecto auth login, or the sign-in ainecto init performs) and never opens a browser on its own. Its role comes only from --role (or AINECTO_ROLE, or the older AI_ERD_ROLE) and is sent with every request; the server applies it.

This is the entry ainecto init writes into a repository's MCP config — you normally do not add it by hand:

{
  "mcpServers": {
    "ainecto": {
      "command": "npx",
      "args": ["-y", "@ainecto/mcp", "--role", "development"]
    }
  }
}

A bridge without a role has no restriction. If a session inside a set-up repository still has no role, another ainecto entry (for example Claude Code's local scope) is usually hiding the one init wrote — check with claude mcp get ainecto. An entry under another name (an older ai-erd one, say) does not hide it but adds a second server with no role — remove it.

MCP Registry

server.json is the Registry entry io.github.codelives/ainecto (same as package.json#mcpName). It lists only the npm package @ainecto/mcp — the per-repository role bridge init writes. Registry publication requires the npm version it references to carry that mcpName, so publish npm first. Registration is done by a person:

1. npm run release -- --otp=<code> (below) so that @ainecto/mcp@<version> is on npm.
2. GitHub org namespace login as a codelives org Owner — a classic PAT with only read:org works where the browser login did not: MCP_GITHUB_TOKEN=<pat> mcp-publisher login github.
3. mcp-publisher publish from this folder. The new version simply replaces the package identifier (@ai-erd/mcp up to 0.4.x → @ainecto/mcp); the server name does not change.

Leave the other entry, com.ai-erd/ai-erd (the remote https://ai-erd.com/mcp, ai-erd.com domain namespace), as it is. One remote URL can belong to one server only; listing https://ainecto.com/mcp as a remote here, or a com.ainecto/* name, would need ainecto.com domain auth (mcp-publisher login dns|http --domain ainecto.com) and is not part of this release.

Development

npm install
npm run typecheck
npm test
npm run build
npm link            # puts `ainecto` (and `ai-erd`) on your PATH

Catalog Sync

sync:tools --check is intended for publish-time live drift checks. It fails fast when AINECTO_CATALOG_SYNC_TOKEN is missing, so automatic publish cannot silently fall back to fixtures. For local development, sync:tools can also use credentials from ainecto auth login --env <env>.

AINECTO_CATALOG_SYNC_TOKEN=... npm run sync:tools -- --env prod --check
AINECTO_CATALOG_SYNC_TOKEN=... npm run sync:tools -- --env dev --check

The checked-in generated catalogs are deterministic output from tools/list, live-synced from https://ainecto.com/mcp and https://dev.ai-erd.com/mcp. Presentation metadata lives separately in src/core/catalog/enrichments.ts. ainecto tools catalog prints the local generated catalog.

Local Tarball Smoke

npm pack
npm exec --package ./ainecto-mcp-<version>.tgz -- ainecto --help

Live Dev MCP Smoke

This requires an authenticated dev token from ainecto auth login --env dev or a local developer AINECTO_TOKEN.

npm run smoke:mcp -- --env dev

The smoke initializes MCP, reads tools/list, calls list_projects, and prints only aggregate metadata such as tool count and Task tool exposure.

Local Attachment Upload Smoke

This exercises the full attachments upload flow against a local API dev server: OAuth MCP token issue, REST fixture document creation, CLI upload, list_attachments, and server filesystem byte verification.

npm run smoke:attachments -- --base-url http://localhost:8080 --endpoint http://localhost:8080/mcp --api-root <path-to-api-repo>

The smoke creates a temporary user/workspace/project/document and removes the fixture account, local temp file, and stored upload file unless --keep is supplied.

Releasing

npm run release -- --otp=<code> runs typecheck, test and build once, then publishes that build under six names: @ainecto/mcp (primary), ainecto, @ainecto/cli (only the ainecto command), and the older @ai-erd/mcp, ai-erd, @ai-erd/cli (only the ai-erd command). --dry-run shows what would go out. An already-published name is skipped, so a retry after an expired code is safe.

Before the first 0.5.0 release, by hand: the npm account needs write access to both the @ainecto and @ai-erd orgs (the script checks), and @ainecto/mcp versions up to 0.3.0 still carry the 2026-09-29 deprecation «moved to @ai-erd/mcp» — replace it, for example npm deprecate "@ainecto/mcp@<=0.3.0" "Old version — use @ainecto/mcp@latest" (new versions are not deprecated). Do not unpublish any name.

Known Limitations

• The MCP connector implements newline-delimited stdio JSON-RPC to HTTP JSON-RPC. Streamable HTTP SSE responses and MCP-Session-Id session handling are not implemented.