---
title: "What works today"
description: "Shipped capabilities, engineering verification and remaining limitations."
canonical: "https://agenttwin.si/docs/status/"
---

# What works today

Shipped capabilities, engineering verification and remaining limitations.


Updated October 22, 2026. The source package and verified npm release are `0.3.14`. The release passed 388 CLI tests across 75 files. Interactive explicit `init` now discovers, connects and syncs, then exits without the control menu; bare `agenttwin` retains the test menu. Node.js 22.12+ is required. Source, published packages and deployed website/API versions must be checked separately. See the [product roadmap](/docs/roadmap/) for release and pilot acceptance gates.

| Area | Available behavior |
| --- | --- |
| Discovery | Static signals, route/job workflow candidates, exercised Node HTTP observations and explicit service confirmation |
| Account sync | Reviewed discovery metadata via authenticated, account-scoped `npx agenttwin sync` |
| Preparation | Explicit application files, invocation, operation mappings, fixtures, faults, assertions and source/spec digests |
| Execution | Local bundled Node execution against HTTP twins by default; optional fresh hosted/E2B sandbox execution |
| Service behavior | Custom JSON collections, bounded Stripe/Slack operations and thirteen additional provider adapter subsets; full vendor parity is not established |
| Infrastructure | Optional cloud PostgreSQL 15 and Redis, with configured final-state checks |
| Reusable worlds | Local authenticated HTTP world host, baseline reset, TTL extension, expiry and teardown |
| Scenarios | Immutable fixture versions, content digests and validated cross-service identities |
| Operation contracts | Explicit supported endpoints, authentication, state effects, errors and limitations |
| Coding-agent tools | Eighteen MCP tools for lifecycle, scenarios, inspection, incidents and protected regression checks |
| Evidence | Assertions, state, reported request call sites, browser step/screenshots, logs and exported repair tasks |
| Dashboard | Real imported/synced resources, movable dotted map, searchable workflows, jobs, cancellation, account menu and saved browser preferences |

See [Discovery, fixes and CI](/stack-loop) for the implemented workflow and exact limits.

## Public agent interfaces

The website now provides a public read-only documentation MCP at `https://agenttwin.si/mcp/docs`, a downloadable skills plugin, an ARD resource catalog, and optional browser WebMCP documentation tools. These serve public guides only. They do not expose customer execution, API OAuth scopes or private shared memory. See the [developer guide](/developers/) for transport versions, request limits and installation.

## Remaining gaps

Arbitrary vendor API compatibility, real third-party account provisioning, native macOS execution, universal workflow extraction and unattended repairs are not implemented. Reviewed local Chromium recipes, bounded background completion, immutable incident regressions and a protected PR check template are available. Protected browser/background and PostgreSQL/Redis CI execution remain unsupported. Coding-agent analysis still needs reviewed invocation, fixtures and outcome expectations.

The API uses a single-instance durable disk store. Distributed scheduling, billing, deletion reconciliation and complete production monitoring remain work to do.

## Verification and data

Engineering checks passed hosted execution, configured PostgreSQL/Redis assertions, account isolation/revocation, cancellation and persistence. The verified October 14 checkout passed 151 CLI tests and 17 web evidence tests. The guided `npx agenttwin init` control menu adds seeded timeout-after-write exploration, captured-source replay, unchanged-check patch verification and saved trial regressions. Seeds control fault selection and delays, not OS scheduling; public npx use requires the updated package to be published. Official Stripe SDK calls, runtime observation, Chromium recipes, background completion, incident imports and CI installation are covered. A real Docker check reproduced duplicate writes, passed the patched commit with unchanged checks, and ignored candidate changes to the grading specification. The packaged MCP loop reproduced duplicate writes, patched only source, reran unchanged checks, verified the receipt and exercised reset, extension and teardown. The October 18 source increment (0.3.8, not yet npm latest when checked October 19) adds thirteen vendored HTTP twins (GitHub, Linear, Notion, HubSpot, Shopify, Datadog, Gmail, Google Calendar, GitLab, Jira, Discord, Twilio, Zendesk), client-side seeded runs with exact `replay`, seed-matrix verdicts, the bounded `shrink` bisection, blackhole/server-error/duplicate-write fault classes, named property checks, and the public docs `v1/ai/search` endpoint; 348 CLI tests and a clean typecheck cover the release. Dashboard checks verified saved preferences, account connect/disconnect and full-width sidebar collapse on desktop and mobile. These checks are not customer adoption evidence or proof of production equivalence.

The production dashboard starts with an empty inventory and receipt list. The separate prototype uses fictional scenarios; its outcomes are scripted.


## Latest integration verification

Main commit `79cbcd3` combines returning-user onboarding and project removal with organization SSO/Keycloak work. Its CLI and docs GitHub checks passed. Combined local verification passed 358 CLI tests (including a browser-suite rerun after installing Chromium), 17 web tests, production builds and browser onboarding/project-removal checks. Commit `3328277` adds shared-context automation with six passing focused tests and a successful GitHub metadata sync. These are recorded engineering results; this documentation update did not rerun those suites. Live OAuth/OTP acceptance, deployment provenance, clean npm onboarding and a complete real pilot remain separate gates.


## Agent integration and website discovery

The website publishes a [developer guide](/developers/), [authentication walkthrough](/auth/), [OpenAPI document](/openapi.json), [API catalog](/.well-known/api-catalog) and [agent documentation index](/llms.txt). The API schema deliberately documents a bounded subset: service health, configured sign-in methods and authenticated account projects. Public documentation does not remove authentication from customer resources.

Coding agents can run `npx agenttwin mcp` over local stdio. Public documentation is available over hosted Streamable HTTP MCP and optional WebMCP browser tools. The product does not currently expose hosted customer-execution MCP, an A2A endpoint, MCP Apps interface, NLWeb streaming service, OAuth authorization server or scoped OAuth account tokens. Adding those requires implemented runtime behavior and security review, not just a discovery file. The private team memory server is separate from the product API.

No fresh external agent-readiness score is claimed here. Directory listings, search ranking and third-party notability require external evidence; website metadata alone does not establish them.