An agent-native platform for business apps. Agents never write application code. They edit a typed, versioned, permissioned model of the app (tables, fields, relations, layouts, workflows, access), and every change they propose is a reviewable diff that a human applies. Most changes wait as drafts until a human publishes them, any published version can be rolled back, and access is enforced per tenant and per row for the people who use the app. Outside agents (Claude Code, Claude Desktop) get the same read and propose tools over MCP: agents propose, humans dispose.
One replayed run of demo-loop.mjs: the 18 steps that CI runs in Chrome on
every code change, with the model's answer read from a recorded cassette, so no model
and no key. 80 seconds; the UI is Vietnamese-first, the captions are English.
Three stills, for readers who skip the GIF
The AI's access card for the sales reps, previewed as ghost cells on the access matrix ("Đề xuất" means proposed).
The rep on the builder's ticket: "this record belongs to someone else: you can view it but not edit it", no Edit.
After the rollback: version 4 republishes version 2; the changes applied at once ("Áp ngay") keep their Revert.
AI app builders that generate code leave one codebase per app to review and maintain. Classic low-code runs a model instead, but was built for a person in a visual editor. Nexus keeps the model and makes it the agent's interface.
| Code-generating app builders | Classic low-code | Nexus | |
|---|---|---|---|
| Who changes the app | an agent, by writing code | a person, in a visual editor | an agent proposes; a person applies |
| What a reviewer reads | a code diff | the change in the editor | a typed diff of one document (a table, a field, a layout, a grant), previewed in place |
| What runs the app | its own generated codebase | the platform's engine | one human-maintained engine for every app |
| Where a runtime fix lands | in each codebase | once, in the engine | once, in the engine |
| Access rules | written and proven per app | the platform's model | tenant_id on every query, entity ACL, row scope |
| What it gives up | a codebase per app to keep healthy | an interface built for agents | whatever the engine cannot express: custom logic is Go work |
The loop and its boundary, as ADR 0002 and ADR 0003 define them:
%%{init: {"flowchart": {"wrappingWidth": 360}}}%%
flowchart TD
S["Builder in the Studio"] -->|prompt| P
M["Claude Code or Desktop,<br/>over MCP"] -->|"tool call"| P
P["propose_* tools validate and buffer;<br/>each proposal is a card: a typed diff,<br/>previewed in place"] -->|"a human applies the card"| K{"Document<br/>class"}
K -->|"Draft<br/>Apply all"| D["Drafts<br/>new schema hidden"]
K -->|"DirectTraced: role, field ACL<br/>Apply now, after a question"| H["Live at once,<br/>with a history entry"]
K -->|"DirectUntraced: relationship…<br/>Apply now, after a question"| N["At once, no history entry;<br/>a new relationship waits<br/>for the next publish"]
D -->|"a human publishes"| V["New version<br/>a rollback is another new version:<br/>newer schema hidden again, nothing dropped"]
V --> R["The customer's users, e.g. the sales rep:<br/>tenant from the token → entity ACL → row scope"]
H --> R
N --> R
R --> DB[("obj_* tables")]
- ADR 0001: the agent's action space is 26 typed tools, 11 that
read and 15
propose_*that validate a change, with the manual builder's validators where one exists, and buffer it. None applies, publishes, writes or takes a tenant. - ADR 0002: one class table decides how an applied change reaches users. Publish is one transaction; a new table or field exists at once but stays hidden behind the schema cutoff. A rollback re-hides newer schema, drops nothing, and leaves applied-at-once changes to Revert.
- ADR 0003: a card follows its class. Drafts go in "Apply all";
access, relationships, a new app and records are "Apply now", after a question.
/mcpserves the same tools to a key that a builder issues, that works on/mcponly and that lasts at most 8 hours. - ADR 0004: records live in one
obj_<entity>table per entity name, shared by tenants; sqlx with allowlisted identifiers there, GORM for fixed-shape tables. - ADR 0005: every query filters
tenant_id, entity ACL runs before the handler and row scope in the repository; the end-user assistant reads through the same checks, as the user who asks. Admins skip ACL, so acceptance runs as the non-admin rep.
docker compose up --buildOpen http://localhost:8080 (set NEXUS_PORT to use another port; NEXUS_SUBNET if its network 10.213.47.0/24
clashes with one you have, and run docker compose down once after changing it). Workspace code: northwind. For the English UI,
open http://localhost:8080/login?lang=en: the choice is kept for the next pages.
| Who | Password | Sees | |
|---|---|---|---|
| Builder (Studio) | admin@example.com |
Demo-Admin-2026! |
everything, edits as drafts |
| Sales rep | test@example.com |
Demo-Rep-2026! |
only what the published version and their role allow: every account, plus their own book: Globex's contact, deal, activity, quote and its lines, contract, and two leads |
These credentials are for your laptop only; docker-compose.yml documents every override. The
builder agent needs an OpenAI-compatible gateway: set AI_ENABLED=true and AI_GATEWAY_BASE_URL, or replay a
recorded cassette as below.
AI_ENABLED=true AI_GATEWAY_MODE=replay AI_GATEWAY_CASSETTE=/cassettes/demo-loop.jsonl docker compose up --build -d
docker wait nexus-seed-1 # prints 0 once the seed has published version 1
(cd nexus-web && npm ci && node scripts/demo-loop.mjs ../demo-loop-shots)
docker compose down -v # a replayed stack answers one runIt prints one ok line per step and saves one screenshot per step; the first failing step exits 1. Set CHROME_PATH
if Chrome is not found, and NEXUS_PORT for both commands if 8080 is taken. For the English run, replay
/cassettes/demo-loop-en.jsonl instead and run the loop with NEXUS_LANG=en: the prompt, the UI and the recorded
answers are English. CI runs both.
Headless Chrome lists only English, and the UI follows a browser that lists English and no Vietnamese. The browser
scripts (demo-loop.mjs, ui-shots.mjs, the others in nexus-web/scripts/) save the language for each page they
open: NEXUS_LANG=vi by default, NEXUS_LANG=en on request. A new puppeteer script gets the English UI unless it
does the same (signIn in studio-driver.mjs).
- In the Studio, open Vận hành › Việc tự động & kết nối › Khoá truy cập and, under Kết nối agent (MCP), click
Tạo khoá cho agent. The key is shown once, lasts at most 8 hours, and comes with the line to run:
claude mcp add --transport http nexus http://localhost:8080/mcp --header "Authorization: Bearer nxm_…" - Ask Claude Code for a change, for example "Add a Suppliers table with name, email, phone and a rating from 1 to 5".
- Back in the Studio, click Mở cuộc trò chuyện on that key: the proposal is a card. Apply it, then publish.
The key acts as the builder who created it and can be revoked from the same list. The server needs no model of its
own for this, so AI_ENABLED can stay off. mcp-loop.mjs runs this loop end to
end, and evals/mcp/ measures it on five requests.
Paths under internal/ are in nexus-platform/.
internal/ddl/ddl.go: the allowlist every table and column name passes before it enters SQL (^[a-z][a-z0-9_]{0,25}$, no reserved words), since identifiers cannot be?parameters.internal/dynrepo/rowscope.go: row scope as a parameterized predicate after the tenant filter: all, team, owner, a user field, shares on reads. An unknown level matches no row.internal/middleware/acl.go: entity ACL before the handler; the HTTP method maps to view, create, edit or delete, denied by default, with one 403 for every refusal. Admins skip it.internal/services/publish_service.go: publish as one transaction under a lock on the tenant row; the reviewed change set, the version and the draft cleanup commit together.internal/metaversion/class.go: the class table, which document types go through drafts, which apply at once with a history entry and which apply at once without one.internal/ai/tools.goandtools_access.go: the builder agent's tools.propose_accessgrants only view, create or edit, only additively, and only on a table proposed in the same turn or not yet published.internal/mcpserver/:/mcp, the same tools filtered to read and propose verbs, run as the builder who issued the key; a proposal lands in that builder's Studio thread.nexus-web/lib/aiApply.ts: how the browser applies a card of each of the 15 kinds, through the same admin APIs as a manual edit. The question before an at-once change is a client rule.
CI runs five jobs on every push to main and every pull request that touches code: Go
(vet, gofmt, test -race, build), Go integration on MySQL 8, web (build, tsc, vitest, eslint), the one-command
demo, and the replayed demo loop in Chrome.
Measured on 2026-10-17 at e0862c1:
| What | Result | Command |
|---|---|---|
| Go unit tests | 3,873 tests (9,603 with subtests) in 42 packages, 0 failed | go test ./... -count=1 -json in nexus-platform/ |
| Go tests on a real MySQL 8 | 4,987 tests (11,215 with subtests) in 49 packages, 0 failed, 0 skipped | bash scripts/run-integration.sh ./... -race -timeout 50m -p 2 in nexus-platform/ |
| Web unit tests | 6,214 tests in 583 files, 0 failed | npx vitest run in nexus-web/ |
| The demo loop, replayed | 18 of 18 steps | node scripts/demo-loop.mjs <dir> in nexus-web/, as above |
| Size | 161,537 lines of Go and 113,083 of TypeScript, tests excluded | wc -l over the tracked non-test *.go, *.ts and *.tsx files |
The MySQL run builds with -tags=integration, so it includes the unit tests: 1,114 of its 4,987 tests exist only
under that tag. run-integration.sh printed PASS=11215 SKIP=0; it
fails any run where a test skips.
- MCP tool-use eval: 5 of 5. Five plain requests to Claude Code connected over MCP, graded by code on the
proposals that reach the Studio; one run each,
claude-opus-5-5, recorded on 2026-10-17 (results). - builder-bench: 20 requests, scored in five stages by code (the right kinds of change; the cards apply; the
version publishes; the gate passes as the non-admin rep; the rep opens the tables).
claude-haiku-5-5: pass@1 0.85, pass@3 0.90 over 60 attempts.claude-opus-5-5: pass@1 0.85 over 20. When the proposal had the right kinds, apply and publish each failed once in 66 attempts, and the rep's checks never failed. Each failure is explained from its transcript; two of the three failing requests are engine gaps, not model errors (report). About 2,000–2,700 output tokens per working app. - Harness eval: regrade WITH 1.00, W/OUT 0.88, Δ +0.12; eval graders WITH 1.00, W/OUT 0.97, Δ +0.03.
claude plugin evalloads.claude/(the builder and verifier agents, the slice and verify-ui skills) as a plugin and runs four small changes to a scaffolded copy of the repo with it and without it: fix a planted tenant-filter leak, add a field with its migration, add an endpoint with its web caller, add a task to the MCP eval. Code graders only, then a regrade that calls the code: hidden behavioural tests, gofmt, vet, the gates from the base commit; a self-test shows that doing nothing, gaming with comments and fourteen wrong-but-plausible changes all fail it.claude-opus-5-5, two runs per case per arm, $12.80, recorded on 2026-10-21. Each delta is one run, and neither is the harness: no with-arm run dispatched its agents or used its skills. One W/OUT run left a filegofmtflags; one correct W/OUT run missed a too-narrow eval pattern (results).
Lights on, not a dark factory (AGENTS.md): one vertical slice per PR, checked as the non-admin rep;
a builder agent writes the tests red first; a verifier agent
with a fresh context tries to refute the PR, and a confirmed HIGH or MEDIUM blocks the merge; every guard is broken
once to watch its test fail. Before merge, the verifier found a privilege-escalation path in
#21 (a proposed table named after a live one let propose_access grant on
the live table), a HIGH in #20 (MCP agents read the published model, not
the builder's drafts, so applying a card would have reverted pending edits), and a demo that contradicted its own
prompt in #16. The long version, with what it cost and what the verifiers
missed: Turning the lights back on.
The threat model is in SECURITY.md: who can do what, the trust boundaries, prompt injection, the
human gate and its limits, and the known gaps. Report a vulnerability privately, through Report a vulnerability
on this repository's Security tab.
-
Runs today: the one-command demo; the metadata engine (tables, fields, relations, layouts, workflows, reports); drafts, publish and rollback; entity ACL and row scope; five installable packs (CRM, which the seed installs, Clinic, Ticketing, LMS, Project Management); the Studio's builder agent and the end-user assistant, which need a model gateway or a recorded cassette; and
/mcp. There are no releases:mainis the product. -
Next, in the open: the engine gaps builder-bench found (#30) and the rest of the issues.
-
Frozen: the packs beyond the demo, approvals (
internal/bpm), work queues, dashboards, calendars, per-tenant backup (internal/tenantbackup) and the Doctor (internal/doctor) are in the tree and tested in CI, but no open issue extends them. -
Vietnamese first, English on the demo path.
?lang=en, or a browser that lists English and no Vietnamese, picks English. The English catalog holds 1,012 of the 6,868 keys innexus-web/lib/i18n: the 421 the demo loop renders (#6 measured them), the sign-in page, the landing page, the assistant panel's keys and its cards, the shell's menus and a few neighbours; other screens fall back to Vietnamese. Still Vietnamese in English mode:- in the assistant: the words of a changed path on an update card, and the approval, work-queue and sample-data
summary lines (
lib/aiBuilder.ts); - from the server or the data: the seeded CRM, the default record layout's section titles, and draft names in the publish list.
The builder answers and labels in the language it is asked in.
cassettes/demo-loop-en.jsonlis that answer in English. The default demo loop, and the GIF above, replaycassettes/demo-loop.jsonl. That cassette was recorded when the builder answered only in Vietnamese, so the names it proposes stay Vietnamese (Phiếu hỗ trợ, Hỗ trợ khách hàng) whatever the UI language. - in the assistant: the words of a changed path on an update card, and the approval, work-queue and sample-data
summary lines (
-
Rollback re-hides; it does not delete (ADR 0002). Schema born after the target is hidden again and its rows stay; attribute edits of a field or table are not reverted; changes applied at once are undone one entry at a time with Revert.
-
Isolation is a convention plus tests, not database-level security (ADR 0005).
-
Open issues that matter:
MIT. Third-party notices: THIRD_PARTY_NOTICES.md.
