tabnas

tabnas/mcp

The tabnas agent tooling: one TypeScript codebase, two front-ends: an MCP server (stdio) and the unified tabnas CLI, sharing a single core (ts/src/core.ts) so the two can never disagree. For each operation the CLI's --json output and the MCP tool result are byte-identical, and the test suite holds them to it.

The website page for this package: per-client setup, the tool contracts, the hosted endpoint's bounds: tabnas.dev/mcp.

Published as @tabnas/mcp. This repo is TypeScript-only: it is tooling over the engine, not a parity package, so there is no Go port.

Install

npm install -g @tabnas/mcp     # the `tabnas` CLInpx --yes @tabnas/mcp          # run the MCP server directly (stdio)

MCP client configuration (stdio):

{  "mcpServers": {    "tabnas": {      "command": "npx",      "args": ["--yes", "@tabnas/mcp@<x.y.z>", "mcp"]    }  }}

Fill in <x.y.z> with the current version: npm view @tabnas/mcp version. This README does not name it: a repo cannot carry an exact pin of its own published version, because the commit that updates it becomes the next release's content, leaving it one release behind forever.

The server is started by the mcp subcommand of the CLI (tabnas mcp), which is exactly what the skills package's mcp.json invokes as npx --yes @tabnas/mcp@<x.y.z> mcp. (--yes matters: on a cache miss npx would otherwise prompt on the stdin the MCP transport owns. Pin an exact version so the tools cannot drift under an installed client: skills/mcp.json carries the real one, written from the registry by its tools/sync-mcp-pin.js and checked by tools/validate.js --online.)

The seven tools

Notes on the contracts:

• Every operation builds a fresh engine instance per call. parse applies options first, then grammar. With no grammar the instance is exactly what new Tabnas() gives: the bare engine defines no rules, so every input yields an undefined tree (serialized as {"ok":true}).
• A grammar argument is validated before it is used, by every operation that accepts one; an invalid grammar is rejected with the validate_grammar error shape {ok:false, errors:[{path,message}]}.
• validate_grammar layers: an ADR-10 security scan (below), structural validation against the bundled grammar.schema.json (Ajv, draft 2020-12), then an engine load in a fresh instance whose thrown message becomes the error. v is the grammar's declared builtin config-schema version (absent means 1).
• Security (ADR-10): a serialized grammar and its options are data, never code. A firewall runs first on every grammar-accepting op and on request options, rejecting: any own key named __proto__, constructor, or prototype anywhere in the tree (prototype-pollution defense, since the engine's grammar merge has no __proto__ guard); a ref key (live functions are not JSON); any function reference that is not a $-suffixed engine builtin; a plugins key, whether a request option or inside grammar.options (a plugin is live code); and grammars over 5000 rules (a CPU bound). "Validate this grammar" never becomes "run this code", or "pollute this process".
• A non-diagnostic engine throw (for example, options.parser.start set to a non-function) is caught and returned as the same clean {ok:false, errors:[{path:"",message}]} shape, so the CLI and the MCP tool agree.
• explain_parse_error joins the diagnostic with the bundled error-code registry entry ({code, message, hint}); registry is null for a code the registry does not know (for example, a plugin-declared code).
• test_grammar takes TSV content in the fleet fixture convention (@tabnas/support): line 1 is a header, the input column is escape-decoded, the expected column is JSON or ERROR / ERROR:<code>. Columns default to positions 0 and 1; options.inputCol / options.expectedCol select by position or header name. Specs over 10000 rows are refused.

MCP resources (served verbatim from the bundled data/): tabnas://schema/grammar, tabnas://schema/diagnostic, tabnas://errors, tabnas://plugins, tabnas://divergence.

CLI

tabnas parse    [file|-] [--grammar g.json] [--json]tabnas validate --grammar g.json [--json]tabnas diagnose [file|-] [--grammar g.json] [--json]tabnas test     --spec fixtures.tsv [--grammar g.json] [--json]tabnas plugins  [name] [--json]tabnas compare  --a old.json --b new.json [--corpus dir|file] [--depth n] [--json]tabnas mcp                                        # run the MCP server (stdio)

Input comes from file, or stdin when the argument is - or absent. The CLI never touches the network. tabnas mcp starts the stdio MCP server (the same server as npx @tabnas/mcp); it speaks JSON-RPC on stdout and prints nothing else there.

--json prints exactly the core result JSON: the same bytes the MCP tool returns for the same request (stable key order; the golden contract, enforced by ts/test/golden.test.js). Without --json you get a readable rendering; a parse failure prints the engine's own rendered error message.

Exit codes:

Passing a grammar that isn't already a GrammarSpec

Every tool takes a serialized GrammarSpec, pure JSON. It does not take ABNF, EBNF, GBNF or jsonic source, and it never will: compiling those means running a compiler, and the tools' one hard rule is that a grammar is data, never code (ADR-10). Compile first, then pass the result.

const { abnfConvert, toPureSpec } = require('@tabnas/abnf')
const spec = toPureSpec(abnfConvert(abnfSource, { builtins: true }))// -> { options, rule, v, meta } — validates clean, safe to send

Use toPureSpec. It is the function for this, and the two obvious alternatives are both wrong:

• abnfCompile() returns jsonic text, not an object. Useful for writing a grammar file; not what a tool argument wants.
• abnfConvert() alone returns a spec carrying ref (empty, when converted with builtins: true) and mark fields. The firewall rejects the presence of ref, not merely its contents, deliberately, since "empty enough" is not a property worth reasoning about at a security boundary, and m marks are not part of the serialized form. toPureSpec strips both and stamps v.

toRecognitionSpec is the same thing for a grammar that only needs to decide accept/reject, without the tree-building builtins.

The equivalent for the other front-ends is tabnas parse --grammar g.json, where g.json is whatever your build step wrote.

Grammar compatibility (compare)

Two questions, reported separately, because they fail differently:

1. Acceptance. Does the candidate still accept what the baseline accepted?
2. Output. For inputs both accept, is the resulting tree the same?

The second is the one users feel. A change that still accepts every historical document but reshapes the tree silently breaks every downstream consumer, and an acceptance-only test reports success.

The report carries evidence and confidence, never a bare verdict. There is deliberately no compatible: true field. Language inclusion is undecidable in general, so a tool that printed one would eventually be wrong in production:

• proven[]. What was established statically, and on what basis. Anything outside the decidable subset is not-proven, which is a statement about this tool, not a claim that the grammars are incompatible.
• observed[]. What actually ran, and how much of it.
• changes[] / counterexamples[]. Concrete differences, with inputs.
• confidence + why. How much weight the absence of findings can bear. confidence: "low" with a stated reason is a successful run.

The check that earns its keep is alternate ordering. Alternates are first-match-wins, so one inserted earlier can shadow a later one and narrow the accepted language while a set comparison calls it an addition. compare walks positions, not membership, and reports a shadowed alternate that used to be reachable.

--corpus takes a .tsv fixture file or a directory of them, loaded through @tabnas/support, the same loader the fixture runners use. Real inputs are the only tier that measures what your documents actually do:

tabnas compare --a v1.json --b v2.json --corpus ../json/test/spec

Exit code is 1 when any change is found, so it works as a release gate.

The hosted endpoint

mcp.tabnas.dev serves the same seven tools over streamable HTTP (POST /mcp, plus GET /health and GET /.well-known/mcp), for agents that cannot run npx. Local stdio stays the recommended path: it is free, private, reproducible, and unlimited.

The hosted service is the same core, so it answers identically; it is also bounded, because it parses attacker-controlled text on shared infrastructure. A 256 KB body cap and 60 requests per minute per IP, both reported up front by /.well-known/mcp and named in the refusal (limit_exceeded / rate_limited) along with the local alternative. The rate limit is Cloudflare's, which counts per IP per data centre and approximately, so it is a shield against sustained abuse, not an exact quota, and a short burst may exceed 60 before refusals begin. Document content is never logged, stored, or used for training; telemetry records shape only: tool name, size bucket, duration, status, error code.

Bundled data

Published packages do not carry the fleet's contract files (the parser's npm files exclude schema/; plugin packages do not ship tabnas.plugin.json), so this repo commits generated copies in data/: the grammar and diagnostic schemas, the error-code registry, DIVERGENCE.md, and every fleet plugin descriptor (plugins.json, sorted by name). Regenerate from sibling checkouts (../<repo> beside this repo) with:

cd ts && npm run gen-data

The build compiles data/ into ts/src/data-bundle.ts (generated, gitignored) and the code reads that static import, never the filesystem, because the hosted Worker does not have one. The test suite fails on a stale regeneration or a stale embed, and checks the embedded set against the directory rather than a hand-kept list. Derive, never duplicate (ADR-10).

data/grammar.schema.json gets the same treatment for a different reason: Ajv validates by generating JavaScript and calling new Function, which Cloudflare Workers forbid outright, so tools/build-validator.js precompiles the schema into ts/src/grammar-validator.js at build time. Same Ajv, same error shapes, compiled earlier.

Build & test

cd tsnpm installnpm test        # builds first, then runs every test/*.test.js

npm test ends with test/workerd.test.js, which boots the real wrangler.json in real workerd and speaks HTTP to it. It is the only test whose failure means "the hosted endpoint will not deploy", and it needs the wrangler devDependency (and the workerd binary npm installs alongside it). It adds a few seconds; a Node-level test cannot replace it, because every deploy-blocking defect this repo has had was green under Node.

Working in the fleet layout (sibling checkouts beside this repo), symlink the siblings after npm install so you test against source (npm replaces these on every install, so re-make them after one):

rm -rf node_modules/@tabnas/parser node_modules/@tabnas/supportln -s ../../../../parser/ts node_modules/@tabnas/parserln -s ../../../../support/ts node_modules/@tabnas/support

CI

.github/workflows/ci.yml, a caller of the org's tabnas/.github polyglot-ci.yml (ts-only, with parser and support cloned as siblings), promoted from ci/ci.yml in a14741c. A change to it is made in .github/workflows/ directly, in a reviewed pull request (admin ADR-8, as amended 2026-09-24).

CI runs test/workerd.test.js, which boots the real wrangler.json in real workerd, so the hosted endpoint's deployability is gated on every push, not discovered at deploy time.

License

MIT. See LICENSE.