Visual Intent Layer
A local-first Visual Direction Loop: open a review surface on an
agent-produced interface, point at visible targets, compose Annotations, and
verify the result by hand. No prose location descriptions, no lost context.
The Annotation is the unit of work. A Builder-Reviewer points at a visible
target, drafts a note in a card anchored to it, attaches reference images, and
queues the result. Sending delivers one Visual Intent Envelope carrying each
Annotation with its own identity. Each Annotation is verified on its own.
The Annotation and the envelope carry one or more targets. Shift extends a
selection into one Annotation's set, and dragging a target already in that set
expresses one relation — ordering, alignment, equal spacing, containment, shared
property or comparative size — shown as one sentence before it is recorded.
The product reviews a browser-rendered surface: a saved or generated HTML
document, or a local running web application. A mobile build, a native desktop
window, a design file and a PDF are outside it by decision rather than omission.
Install
Requires Node 20+. No Rust toolchain, no hosted account.
The product is a standard local MCP server. You register it with your own agent's
MCP mechanism — there is no install command of ours to run, and nothing of yours
that we write.
| Claude Code | claude mcp add visual-intent-layer -- npx -y --package visual-intent-layer@latest visual-intent-mcp |
| Codex | codex mcp add visual-intent-layer -- npx -y --package visual-intent-layer@latest visual-intent-mcp |
| opencode | opencode mcp add visual-intent-layer, then give it npx and the arguments above when it asks |
| pi | with pi-mcp-adapter installed, the same entry in the project .mcp.json, or the adapter's /mcp panel |
| any other | any host that can launch a local stdio server: command npx, arguments -y --package visual-intent-layer@latest visual-intent-mcp |
If you installed the package globally, the command is simply visual-intent mcp.
Registry-aware clients can also find the server as
io.github.willa-code/visual-intent-layer in the official MCP Registry, which lists
where to get it rather than hosting it.
Install it on the machine where you look at the screen. The Review Surface
launches a browser on the machine running the server and binds a loopback port
there. A server in a container, on a remote host, or inside a cloud agent has
nothing to show you.
Registering takes effect when your agent next starts, so restart it. If the tools
do not appear, your agent's own MCP panel is where the reason is: it names the
server and reports why it failed to start. The product cannot tell you — it writes
no configuration and reads none.
The command carries @latest, so each session resolves the current release and
nothing needs updating by hand. Pin an exact version instead if you would rather
nothing move under you.
The same server, without an agent
The local commands need no agent and no registration:
visual-intent open --html ./checkout.html
visual-intent open --app http://localhost:5173
visual-intent serve --port 3742
This is the route for driving the Review Surface yourself, and the fallback when a
running service must be started by hand. It opens a review but delivers nothing: with
no agent holding the loop, an envelope has nobody to reach.
Use
The loop starts when an agent calls open_visual_review, or when you start a session
yourself — the three commands are in The same server, without an agent.
Starting a session launches the default browser on this machine and prints
the review URL as well, so the loop is a product experience rather than a
URL to copy. Automatic opening is suppressed with --no-open or
VISUAL_INTENT_NO_OPEN=1 for headless, remote and scripted use; the URL is
still printed. When no browser can be opened on this machine at all, the product
says so and prints the URL rather than failing silently, and the agent is told the
same in the tool result. Reopening the same artifact revision reuses the open
session and its URL, so a review does not accumulate tabs.
One rail, two tiles
The Review Surface has one chrome region and one state. The rail on the
right holds the artifact's identity, its revision, the agent's position and
every Annotation — unsent and sent alike, in one list with the state pill
carrying the difference, so nothing leaves the view when it is sent. An
Annotation that needs a decision comes first; closed ones stay behind one
toggle. Verdicts are decided where the Annotation sits: approve, reject with
another pass, or mark obsolete. Amending one that was already sent supersedes
it and delivers the amendment; it is never rewritten in place.
A small island over the artifact holds two icon-only tiles: point at things,
and box an area. Operating the artifact is the unarmed resting state rather than
a third tile. Pointing is a gesture: clicking targets a thing the artifact owns,
and dragging across words targets exactly those words. Shift extends the
selection into a set of up to eight targets and removes a member already in it;
a plain click or box replaces the set. Dragging a target that is already in the
set is a relation drag: it moves a ghost and states one sentence before anything
is recorded. A drag that begins anywhere else behaves as the artifact does, so
text selection is unaffected. P, B and V arm and disarm the tiles.
Type and press Enter in the Annotation card to queue it; Cmd/Ctrl+Enter sends
the whole queue. Escape unwinds exactly one level — close the card, then clear
the selection, then return to operating the artifact — and never discards unsent
writing.
The Annotation model
- An Annotation is durable and individually identified: targets, note,
references, relations, delivery state and resolution. A relation is stored as
the desired relationship, with no pixel field, and the agent chooses how to
achieve it.
- Unsent text survives a surface reload, a service restart and a browser
restart. Nothing discards a note silently.
- An Annotation carries a set of targets.
Shift extends the set or removes a
member; a drawn Area is one member and may take part in a relation. A relation
is shown as one sentence in the card and in the rail row, from one
implementation.
- Every Target carries Runtime State Evidence: the address the artifact was
showing when it was pointed at, recorded relative to the artifact's own base
so it means the same thing whether the artifact is served directly or through
the review proxy.
- A Target may carry a Captured View: a browser-composited image of the artifact
as it was seen, taken by an explicit permissioned capture in the tab that
shows it and never substituted by a re-render. It is content-addressed like a
reference image and left undisclosed to no one: the drawer says it will leave
the machine before the queue is sent.
- Reference images are added by picker, paste or drop, are content-addressed by
their own bytes, and are refused visibly (and unread) when disallowed or
larger than 5MB. Only image types are accepted.
- Relational Intent is expressed by manipulating targets directly:
Shift
extends a selection into a set, and dragging a member of that set infers one
relation, shown as one sentence before it is recorded. The relation is stored
in the envelope as the desired relationship, with no pixel value anywhere, and
the artifact's source and authoritative DOM are never changed by the drag.
- The product never writes style values into the artifact or its source.
Resolution, honesty and the agent
When the artifact changes, each target is located again and reported as
Matched, Recovered, Ambiguous (with its candidates, never
auto-selected) or Deleted. A target that cannot be found while the
annotation's revision is still the one on screen, and that was pointed at a
different address than the one now showing, is reported as possibly existing
only in a state no longer on screen — the product states that possibility and
never asserts it. Whether an Annotation was written before the revision now on
screen is a separate, Annotation-level fact shown as such. Provenance
Confidence — exact source span, inferred, or unavailable — is a separate axis
and never shares the word "exact" with target resolution; exact is claimed
only where an instrumented artifact stamped a source location, per Target — a
stamp read from data-vis-source, or from data-insp-path written by
code-inspector-plugin as one way to produce it.
Every Annotation is stamped with the Adopted Revision: the revision the
artifact reported it was holding, not the revision a source currently offers.
Where the artifact's report and the source's offer disagree, the surface states
the disagreement and the Annotation keeps the artifact's report.
The surface states the agent's position in a sentence, whether a tool call is
currently held or direction will be read at the agent's next Check-In, and when
the agent last checked in. Where the host can hold the call the agent is
awaiting you and sending is urgent; where it cannot, the surface says the
agent has stepped away and the Annotations are queued durably. Agent
acknowledgement is never presented as implementation or as verification, and the
workflow never requires the agent to be waiting.
Check-In is a convention, not a capability
Steering and interruption are seen at a Check-In: the point between an
agent's own steps where it reads new direction. There is no push channel and no
wake mechanism, because a server cannot put anything into an agent's running
turn. The product publishes the convention instead of detecting a host
capability it cannot rely on, and records every MCP call for a session so "last
checked in" is a fact the surface states rather than an assumption.
An agent that was not holding a call can retrieve what arrived with the
check_in tool: newly delivered Annotations with their intent, amendments that
replaced something, a pending stop request, and the current state of what it
was given before. Sending the queue is Next-Pass Intent, amending something
already sent is Steering Intent, and asking an agent to stop is Review
Interruption — a request, never a claim that work stopped. The envelope's
review-interruption value is reserved and never emitted: an interruption names
no target, so it cannot be one.
The nearest canvas-agent prior art refuses check-in outright and schedules later
requests instead; polling through the MCP Tasks extension is the one push-free
pattern the specification sanctions. Both are recorded here.
A single drawer, hidden while it has nothing to report and badge-counted when it
does, carries what needs a decision: everything that will leave the machine, any
unresolved or ambiguous Annotation, and the fact that the artifact has moved on.
The overflow menu holds the rare actions — reload artifact, copy artifact path,
copy evidence for the queue, open the disclosure, end session, and choose the
chrome theme — and never a frequent one.
Agent tools (MCP)
The model-visible tool surface is deliberately small:
open_visual_review | Opens the loop for a saved HTML file or a running localhost app, may declare capabilities: { embeddedUI, subscriptions }, and holds the call for the human. |
check_in | Reads new direction between the agent's own steps, without an envelopeId: deliveries with their intent, amendments as replacedId/replacementId pairs, a pending stop request, and current state. |
get_intent_status | Reads one delivered batch by envelopeId, including each target's resolution. |
acknowledge_intent | Confirms receipt of a batch or one Annotation. Never implementation and never verification. |
embeddedUI and subscriptions are the only host capabilities the product
negotiates; there is no steering flag. A host declares them through the
capabilities argument on open_visual_review, not through a side channel.
The tool descriptions ship the Check-In convention, so an agent following them
knows to call check_in between its own steps.
Environment and data
VISUAL_INTENT_DATA_DIR | Lifecycle data directory (default ~/.visual-intent-layer/data). |
VISUAL_INTENT_PORT | Default service port (default 3742). |
VISUAL_INTENT_NO_OPEN | Set to 1 to suppress automatic browser opening. |
VISUAL_INTENT_WAIT_MS | How long the agent-facing entry tool holds the call, in milliseconds. |
Annotations, sessions, attachments, snapshot bytes and Check-In contact all live
under the data directory, so a service restart loses nothing. The packaged
Envelope schema
The portable contract is schema/envelope-v0.4.schema.json (experimental,
versioned). One envelope carries one or more Annotations, each with its own
identity, targets, note, relationships, references and attachments. Every target
may carry Runtime State Evidence, including the address the artifact was showing
and, for a Target reached through a frame, the ordered chain of documents it was
reached through. TypeScript types are generated from it (npm run build:types).
schema/envelope-v0.1.schema.json is kept only as a historical record of the wire
shape, including the retired supersedes field name: validateEnvelope reads 0.4,
0.3 and 0.2, and an envelope shaped like 0.1 is refused. Store files written
against the older Annotation model are migrated once, at store startup, by
src/annotation/migrate.ts, independently of schema-version negotiation. A 0.2 or
0.3 envelope loads unchanged and is read as the current version. New envelopes are
0.4.
review-interruption remains a reserved value in the delivery.intent enum and
is never emitted, because an interruption names no target and so cannot be an
envelope.
Security and privacy
Reviewing is local-first. The service binds to loopback only, requires an
unguessable per-session capability, and confines file access to the artifact
directory. A saved HTML artifact keeps its own relative and root-relative assets
and may load the remote stylesheet, font and image origins it declares — and
those origins are disclosed in the surface before the artifact contacts them.
Runtime data requests stay blocked by a saved artifact's content policy. A
running local development server may be reverse-proxied through the review
service's own origin so its DOM is selectable; only loopback development origins
are proxied, a client's method, body, headers and redirects arrive intact, its
update channel is proxied, and authenticated production applications are
refused. A proxied application gets its own content policy that permits its own
proxied origin and nothing else, and the disclosure states what is permitted. A
Captured View is taken only by an explicit permissioned capture in the
reviewing tab; the permission cannot be persisted.
See SECURITY.md for the full policy.
Develop
npm install
npm test
npm run typecheck
npm run build
npm run benchmark
npm run instrument
npm run eval:invocation
The primary test seam is tests/browser-loop.test.ts: the Visual Direction
Loop is driven end to end in a real browser engine against the locally running
service — the served asset graph must load, the artifact must render with its
styles, and a human path (point, annotate, queue, send, reload, compare, amend,
stop, verify, restart) is performed through the DOM the product actually serves.
A surface that cannot load its own scripts fails CI rather than shipping.
The design gallery is at /gallery on a running service. It is not part of the
product's navigation. tests/gallery-snapshot.test.ts pins every design token
exactly, fails on a screenshot difference, checks every semantic surface/ink
pair against the contrast floor in both themes, and fails when a baseline is
missing rather than silently writing one. Tokens are pinned once because they
are strings; screenshots are pinned per rendering platform, because font
metrics move the layout — a missing platform baseline fails and leaves the
render this run produced in .scratch/, which CI uploads as the
gallery-actual artifact. Regenerate baselines for the platform you are on
with UPDATE_GALLERY=1 once the change is intended.
Layout
VISION.md — the human-in-the-loop thesis and the UI/UX standard every decision defers to
design.md — normative Review Surface design contract (tokens, roles, states)
CONTEXT.md — the domain language, including the words to avoid
schema/ — versioned Visual Intent Envelope contracts (0.4 current; 0.2 and 0.3 read for compatibility; 0.1 kept as a historical reference)
src/annotation/ — the Annotation model, durable store, attachments, migration
src/artifact/ — identity, content-addressed revisions, fidelity rewriting, snapshots
src/resolution/ — target resolution and its reduced vocabulary
src/mcp/ — MCP server, the four tools, stdio transport
src/service/ — loopback HTTP service, sessions, Check-In records, security boundary, browser opening
src/ui/ — product-owned shell, artifact interaction layer, icon set, design gallery
src/adapters/ — the build-time source-location stamp reader (product and third-party attributes)
src/host/ — the surviving host-capability declaration (embedded UI, subscriptions)
src/benchmark/ — mutation benchmark matrix
src/instrumentation/ — product-boundary measurements
src/eval/ — documented invocation-routing eval
fixtures/ — the gallery artifact the Lever and tests drive
tests/ — the browser loop, the Lever contract, and the gallery baselines
scripts/ — UI bundling and envelope type generation
Documentation
docs/adr/ — architecture decisions, including ADR-0018 on Check-In replacing
a steering capability
docs/background/ — research and the product strategy brief
docs/pi-validation.md — the live-pi checklist
docs/agents/ — the issue tracker, triage labels and domain-doc conventions
SECURITY.md — the threat model, supported versions and disclosure process
Maintaining this package
Releases go out through the Publish workflow (npm trusted publishing).
/maintain-visual-intent-layer lays out the paths — dogfood locally, cut a release,
or fix a bad release — and executes the one you pick. Every published version is an
official release and latest always names the newest of them.
License
Apache-2.0. The complete local loop is permissively open; see SECURITY.md
for the threat model and disclosure process.