Request access

Private betaNow watching TypeScript repositories on GitHub

Your code didn't change.The API did.

Mendward watches the APIs and SDKs your code depends on, proves which lines a change breaks, writes the smallest safe fix, runs your own checks, and opens a pull request with the evidence.

Request access

The upstream example on this page, a Ledger API renaming entryRef to entry_id, is synthetic. It mirrors a real benchmark case.

How an upstream change breaks production silently

  1. It compiles.
  2. Tests pass.
  3. Production breaks.
  4. A customer finds it first.

Nothing in your repository changed, so nothing in your pipeline looked. The provider renamed one response field on a Tuesday.

Every provider.Every version.Every call site.

Nobody reads them all.

Mendward does, and only tells you about the ones that reach your code.

Sixty-eight seconds, from the crack to the mend.

The launch film. Every frame is a function of time, every sound is synthesised, and every hit lands on the beat.

Detect the change. Prove the impact. Ship a verified fix.

Six stages, in a fixed order, recorded as data. Each one hands the next a fact, never a guess.

  1. 01

    Ingest

    Fetch the upstream source and keep the exact bytes.

    Mendward fetches the published specification, SDK release or schema, stores the raw bytes with a SHA-256, and normalises it. The fingerprint is the only identity a fetch has.

  2. 02

    Detect

    Diff it, and explain why each change matters.

    A semantic diff against the last observation produces typed changes with evidence: what moved, where in the document, how severe, and the reason for that severity.

  3. 03

    Analyze

    Prove which lines of your code it reaches.

    Mendward builds a real program from your repository and follows symbols through imports, aliases, wrappers and re-exports. Every finding carries a file, a line, a column and the evidence for it.

  4. 04

    Migrate

    Write the smallest safe patch.

    Deterministic recipes run first. A model is used only for what no recipe claims, and its hunks are labelled. Every changed line must be attributable to a specific edit, or the patch is refused.

  5. 05

    Verify

    Run your own checks in a sealed container.

    Your typecheck, build and tests, as your repository declares them, in a container with no network and a read-only worktree. A check that did not run says so, forever.

  6. 06

    Pull request

    Open a pull request a reviewer can trust.

    The diff, the evidence, the verification results, what to doubt and how to undo it. One idempotency key per change set, so a retry can never open a duplicate. You merge. Mendward never does.

01

Ingest

Mendward fetches the published specification, SDK release or schema, stores the raw bytes with a SHA-256, and normalises it. The fingerprint is the only identity a fetch has.

ledger-api / openapi.jsonObserved
Fetched
2026-10-05 09:14:02 UTC
HTTP
200, 412 KB, ETag "7c1e"
SHA-256
9f3c4a1e...b07de21a
Contract
OpenAPI 3.0, 86 operations, 140 schemas
Previous
41aa09c2...5e1f7730

The fingerprint is the fetch's only identity. Same bytes, no new run.

02

Detect

A semantic diff against the last observation produces typed changes with evidence: what moved, where in the document, how severe, and the reason for that severity.

Semantic changes2 found
property_renamedBreaking

LedgerEntry.entryRef to LedgerEntry.entry_id

A response property your code may read no longer exists under its old name. Direction: response.

#/components/schemas/LedgerEntry/properties

property_addedAdditive

LedgerEntry.posted_at

New optional field. Nothing to migrate.

03

Analyze

Mendward builds a real program from your repository and follows symbols through imports, aliases, wrappers and re-exports. Every finding carries a file, a line, a column and the evidence for it.

Impact on acme/payments-service3 findings
src/ledger.ts:42:21HIGH
const ref = row.entryRef;
  • Resolved by the type checker to LedgerEntry
  • Import traced to @ledger/sdk
src/report/export.ts:18:9HIGH
const { entryRef } = entry;
  • Destructured read of the renamed field
scripts/legacy-fmt.ts:7:14LOW
log("entryRef missing");
  • Text match only. A string that contains the name is not proof it is read.
04

Migrate

Deterministic recipes run first. A model is used only for what no recipe claims, and its hunks are labelled. Every changed line must be attributable to a specific edit, or the patch is refused.

src/ledger.tsrename_property, deterministic
41export function toRow(row: LedgerEntry) {
42  const ref = row.entryRef;42  const ref = row.entry_id;43  return { ref, amount: row.amount };
44}
Every changed line attributed to an edit2 of 2
05

Verify

Your typecheck, build and tests, as your repository declares them, in a container with no network and a read-only worktree. A check that did not run says so, forever.

Verification, no network, read-only /work
✓pnpm install --frozen-lockfile41.2s
✓pnpm typecheck12.8s
✓pnpm build33.0s
✓pnpm test214 passed
!lintSKIPPED_WITH_REASONNo lint script is declared in package.json. Reported as skipped, never as passed.
06

Pull request

The diff, the evidence, the verification results, what to doubt and how to undo it. One idempotency key per change set, so a retry can never open a duplicate. You merge. Mendward never does.

acme/payments-serviceOpen

Migrate Ledger API: entryRef renamed to entry_id #128

mendward/ledger-entry-id into main. 2 files, +2 −2.

  1. Upstream change
  2. Why this repository
  3. What changed
  4. Verification: 4 ran, 1 skipped with reason
  5. Known uncertainty and manual review
  6. How to undo this
  7. Evidence and confidence: HIGH

Mendward opens it. You merge it.

No affected code. No pull request.

A tool that opens forty pull requests gets uninstalled. Mendward opens one where your code reads the thing that changed, and stays silent everywhere else.

acme/payments-servicePull request #128
  • src/ledger.tsreads entryRef
  • src/report/export.tsreads entryRef
  • src/api/client.ts
  • src/routes/entries.ts
  • test/ledger.test.ts
  • package.json
acme/marketing-siteNo pull request
  • src/pages/index.tsx
  • src/components/Hero.tsx
  • src/lib/analytics.ts
  • src/styles/site.css
  • next.config.ts
  • package.json

Nothing here reads entryRef. Checked, recorded, and not your problem.

No percentages. A level, and the signals behind it.

A percentage reads like a measurement nobody took. Mendward reports LOW, MEDIUM or HIGH, computed two ways and the lower one wins: a cap from the worst negative signal, a raise from the best positive one. Try it.

src/ledger.ts:42:21Try the rule
Evidence for the finding
Evidence against
RaiseHIGH
CapHIGH
ReportedHIGH

The lower of the two wins. Two independent signals put it at HIGH, and nothing argues against it.

Not a vendor list.

A customer should be able to point Mendward at an API it has never heard of. Every source is compiled into one contract model, and everything after that is shared, so a new company never needs new code.

How any source becomes a pull request
  • OpenAPILive
  • npmIn development
  • PyPIIn development
  • GraphQLIn development
  • ProtobufIn development
  • JSON SchemaIn development
  • GitHub releasesIn development
  • ChangelogIn development
  • Your webhookIn development

One contract model

Provider-neutral. The diff never learns who published it.

  1. Semantic changes
  2. Impact
  3. Migration
  4. Verified pull request
OpenAIStripeTwilioSupabaseGitHubAnthropicAWSGoogle CloudShopifySlackSendGridCloudflareFirebaseVercelNetlifyAuth0ClerkOpenAIStripeTwilioSupabaseGitHubAnthropicAWSGoogle CloudShopifySlackSendGridCloudflareFirebaseVercelNetlifyAuth0Clerk

Live today Preset in development. Names are plain text: no logos, no implied partnership.

Built to be trusted with private code.

Mendward will hold some of the most sensitive repositories a company has. Security is not a feature list here. It is a set of properties, each one proven by a test that fails if it stops being true.

  • Four GitHub permissions. contents read, contents write, pull requests write, metadata read. Everything else is on a forbidden list that fails the build.
  • Tokens are never stored. Installation tokens are minted on demand, cached in memory for at most an hour, and never written to a database.
  • Tenants cannot see each other. Row-level security is enabled and forced on every tenant table, and a cross-tenant request gets the same 404 as a real miss.
  • Your code never reaches a prompt as instructions. Customer source and committed secrets are kept out of model requests, proven by 45 tests on that boundary.
Read the security model
docker runasserted by a golden test
  • --rm --initgone when it exits
  • --read-onlythe root filesystem cannot be written
  • --cap-drop ALLno Linux capabilities at all
  • --security-opt no-new-privilegesnothing can escalate
  • --network noneno way out
  • --pids-limit --memory --cpushard ceilings, swap included
  • --ulimit core=0no secret-bearing core dumps
  • --mount /work,readonlyyour code, read-only
  • --env NAME=valueexplicit values, never inherited

Twelve things Mendward will not do.

These are the product. A change that breaks one is a regression, whatever else it improves.

  1. 1

    It will not report a check as passed when it did not run.

    A skipped check is recorded as skipped, with the reason, in the database, the interface and the pull request. It never turns green later.

  2. 2

    It will not output a confidence percentage.

    You get a level, LOW, MEDIUM or HIGH, plus every signal that produced it. Three weak findings and one strong one do not average to 72%.

  3. 3

    It will not fabricate a number.

    Counts, lines, durations and outputs come from a measurement a code path performed. An empty value means nothing was measured, never zero.

  4. 4

    It will not touch a line no edit claimed.

    Every changed line is attributed to a specific edit and checked against an independently computed diff. Reflowed whitespace is refused outright.

  5. 5

    It will not hide a repository it could not verify.

    If verification could not run, the pull request says so on its first screen, not in a footnote.

  6. 6

    It will not read your repository as instructions.

    A model sees repository text only in a delimited data channel, with a static system prompt, provenance on every fragment and visible truncation.

  7. 7

    It will not ask GitHub for a permission it cannot justify.

    Four permissions, each traced to one product behaviour. Anything it might be tempted to add is named in a forbidden list that fails the build.

  8. 8

    It will not run your code outside a sealed container.

    No network, a read-only worktree, dropped capabilities and hard resource ceilings. Production refuses to start on an unpatched runtime.

  9. 9

    It will not forward a secret into a container, an error or a log.

    Environment is passed by explicit value only, and 23 control-plane secret names are refused even when asked for.

  10. 10

    It will not delete or edit an audit row.

    The audit trail is append-only at the database level and hash-chained per organization, so an edit breaks the chain.

  11. 11

    It will not tell you your code is safe when it does not know.

    An empty set of findings is a hard failure, not a reassuring empty plan. Not knowing is reported as not knowing.

  12. 12

    It will not average.

    If four places need a human, it says four places need a human, and names them.

  13. All twelve, with the reasoning behind each.

    Read the commitments

Every number on this page came from a run.

Mendward is built the way it behaves. The release gate tests itself before it tests anything else, and refuses to start if its own pass and fail logic is wrong.

Unit tests
Integration tests
Browser tests

Measured 2026-10-04 at commit 596203d, against real Postgres and a real container runtime.

node scripts/verify.mjs
  1. PASSlint0 errors5.5s
  2. PASStypecheck:packagesclean0.8s
  3. PASStypecheck:webclean25.7s
  4. PASStest:gate50 passed11.6s
  5. PASStest:unit2,657 passed in 93 files60.1s
  6. PASSdb:verify20 tables, 8 functions, 0 violations1.8s
  7. PASStest:integration543 passed in 36 files266.7s
  8. PASSbuild:webproduction build58.3s
  9. PASStest:e2e20 Playwright specs35.6s

9 passed, 0 failed, 0 skipped, 9 total

Asked, and answered plainly.

The short version of what Mendward is, what it touches, and what it will not do. The pages below show the detail.

How it works What is covered Security
What is Mendward?
Mendward is autonomous software maintenance for code that depends on third-party APIs and SDKs. It watches those upstream sources, detects a breaking change, proves which lines of your code it reaches, writes the smallest safe fix, runs your own checks in a sealed container and opens a pull request with the evidence.
What does it support today?
In the private beta, Mendward works on TypeScript repositories on GitHub, with OpenAPI as the live contract format and presets for OpenAI, Stripe, Twilio and Supabase. Every other provider, format, ecosystem and language is listed on the coverage page with the exact level it has reached.
How is it different from Dependabot or Renovate?
Those tools raise the version of a dependency. Mendward starts from what changed in the API itself, finds the code that change reaches, edits that code, and runs your checks before it opens a pull request. It also watches HTTP APIs that are not packages at all.
Does Mendward use AI?
Deterministic first. Detection, impact analysis and recipe edits are computed without a model. A model is used only for what no recipe claims, sees repository text only in a delimited data channel, and its hunks are labelled in the pull request. No model is required to use Mendward.
What access does it need to my GitHub repositories?
Mendward is a GitHub App with four permissions: contents read, contents write, pull requests write and metadata read. Installation tokens are minted on demand, kept in memory for at most an hour and never written to a database.
How confident is a fix?
Every finding carries a level, LOW, MEDIUM or HIGH, never a percentage. It is computed two ways and the lower one wins: a cap from the worst signal against the finding, and a raise from the best signal for it. Every signal is shown.
Does it merge anything by itself?
No. Mendward opens a pull request for a person to review. If a check did not run, the pull request says so on its first screen, and a skipped check is never reported as passed.
What does it cost?
Four plans, Indie, Startup, Growth and Enterprise, shaped by repositories, members, migration runs and verification minutes, with a 14-day trial. Prices are published at general availability. During the private beta, terms are agreed with each team.
0:00 / 1:08
Esc
UpDown to moveEnter to runType / for a few more