MCP

The Lace MCP server connects your agents to everything your team captures. Comments carry the screenshot, the spot on screen, and the discussion behind them. Resolving a comment promotes it to a decision. Recorded sessions expose the ordered screens and actions from capture. Agents search your comments and sessions, pull the full context, and execute: they can rerun a workflow the way you ran it, or take on follow-up work with your reasoning in hand.

Connect in Claude

The fastest way to use Lace is the hosted connector at https://mcp.inlace.co. It runs in the cloud and reads your comments and decisions straight from your Lace account, so you don’t need the desktop app open.

Claude.ai
Claude Code

Open Settings → Connectors → Add custom connector.
Enter the URL https://mcp.inlace.co and click Add.
Click Connect and sign in with your Lace account.

Then start a chat and say “Search Lace decisions”.

claude mcp add --transport http lace https://mcp.inlace.co

The first call opens a browser to sign in with your Lace account. After that, say “Search Lace decisions”.

The hosted connector needs a Lace account and at least one resolved comment. Sign-in uses OAuth 2.0, so the first connection opens your browser.

Install from the CLI

Use the Lace CLI to register the right MCP transport for each editor. Claude Code uses the hosted OAuth connector so it can manage authentication in /mcp; Claude Desktop, Cursor, and Codex use the local stdio server for editor-side installs.

Cursor
Claude Code
Claude Desktop
Codex CLI

Or install from the terminal:

npx lace-mcp install cursor

Restart Cursor, then open a new chat and ask about resolved Lace decisions.

npx lace-mcp install claude

Start a new Claude Code session, run /mcp, and authenticate Lace. Claude Code uses the hosted OAuth connector so its MCP menu can show tools, re-authenticate, clear authentication, reconnect, and disable.

npx lace-mcp install claude-desktop

Restart Claude Desktop and say “Search Lace decisions”.

npx lace-mcp install codex

Start a new Codex session and say “Search Lace decisions”.

Claude Desktop, Cursor, and Codex local installs only work while the Lace desktop app is open and you’re signed in. The Claude Code hosted connector at mcp.inlace.co does not need the app open.

How it works

Write a comment

In Reviews, write a comment and attach screen context — a pin on the exact spot you’re looking at. Resolve it into a decision. The screenshot, spatial context, and discussion persist with it.

Your agent queries Lace

It calls search_comments to find relevant comments, picks one, then calls apply_comment for screenshots, element metadata, and step-by-step context. For what you did on screen earlier, it calls search_sessions then get_session_flow.

Your agent executes

It carries out the work and you review the result.

Technical reference

MCP server

The Lace MCP server is registered under the name lace. It exposes comment/decision tools, session-flow tools, and publishing.

Session tools

Session data resolves in tiers — prefer the cheapest tier that answers: search_sessions (find) → get_session_flow (skeleton) → include: "elements" for specific screens → include: "screenshots" last. Sessions you create are available to you automatically. Another user’s sessions appear only when that user has turned on sharing for the session.

`search_sessions`

Find recorded screen sessions — episodes of what the user did on their machine. Returns up to 20 summaries: session id, time span, apps, screen/step counts. Never returns steps, elements, or images. Parameters:

Parameter	Type	Default	Description
`query`	string	—	Free text over window titles and on-screen element/target labels
`appName`	string	—	App name filter
`url`	string	—	URL path filter
`startedFrom`	string (ISO date)	—	Lower bound for session start
`startedTo`	string (ISO date)	—	Upper bound for session start
`limit`	number	`20`	Max summaries to return
`cursor`	string	—	Opaque pagination cursor from a previous response

All parameters are optional. Returns: session summaries. Next step: get_session_flow with a returned sessionId.

`get_session_flow`

Replay one recorded session as an ordered flow. Parameters:

Parameter	Type	Default	Description
`sessionId`	string (UUID)	—	Session ID from `search_sessions` (required)
`include`	`elements` \| `screenshots`	omitted	Optional detail tier (see below); omit for skeleton
`screenKeys`	string[]	—	Required for `elements` and `screenshots` — which screens to expand

include options:

Omitted — ordered skeleton: actions, targets, dwell, and screen identity (app, window, url, screenKey). No elements, no images. Capped at ~200 steps.
elements — CapturedElement trees (role, label, bbox) for the listed screenKeys only. Use this to reason about what was on screen.
screenshots — at most one inline image plus download links for the listed screenKeys. Use only for visual/layout questions the element tree cannot answer. Never request screenshots to identify or locate elements.

Returns: the requested tier, plus a plain-text hint for the next cheaper/richer call.

Discovery tools

`search_comments`

Search and filter comments with cursor pagination. With no parameters, returns the top 20 resolved comments ranked by recency. Use the filter parameters to narrow results by review, project, app, element, time range, and more. Parameters:

Parameter	Type	Default	Description
`query`	string	—	Text search across headline, body, and element text
`status`	`resolved` \| `suggested` \| `all`	`resolved`	Filter by comment status
`limit`	number (1–100)	`20`	Results per page
`cursor`	string	—	Pagination cursor from a previous response
`sort`	`relevance` \| `resolved_at_desc` \| `created_at_desc`	`resolved_at_desc`	Sort order
`reviewId`	string (UUID)	—	Exact review ID
`reviewTitle`	string	—	Review title substring match
`reviewStatus`	`active` \| `archived`	—	Review lifecycle status
`pageId`	string (UUID)	—	Exact page ID
`pageName`	string	—	Page name substring match
`projectId`	string (UUID)	—	Exact project ID
`projectName`	string	—	Project name substring match
`threadId`	string (UUID)	—	Filter by thread ID
`appName`	string	—	App name substring match
`windowTitle`	string	—	Window title substring match
`url`	string	—	URL substring match
`anchorState`	`anchored` \| `reanchored` \| `detached`	—	Anchor state filter
`elementEid`	string	—	Exact element accessibility ID
`elementRole`	string	—	Exact element ARIA role
`elementInteractivity`	boolean	—	Filter to interactive elements only
`minEndorsements`	number	—	Minimum endorsement count
`hasReplies`	boolean	—	Filter to comments with replies
`createdBy`	string (UUID)	—	Creator user ID
`resolvedFrom`	string (ISO date)	—	Lower bound for resolved date
`resolvedTo`	string (ISO date)	—	Upper bound for resolved date
`createdFrom`	string (ISO date)	—	Lower bound for created date
`createdTo`	string (ISO date)	—	Upper bound for created date
`compact`	boolean	`false`	Return summary fields only to reduce token usage when scanning

All parameters are optional. Returns: a paginated list of comment summaries including headline, element info, review/project context, timestamps, and a nextCursor for the next page.

`list_reviews`

List reviews in the current org. Returns title, project, page names, and comment counts per page. Use review titles and IDs to scope search_comments queries. Parameters:

Parameter	Type	Default	Description
`status`	`active` \| `archived`	`active`	Review status filter
`content_type`	`review` \| `artifact`	—	Filter by review content type
`projectId`	string (UUID)	—	Filter by project ID
`limit`	number (1–100)	`50`	Results per page
`cursor`	string	—	Pagination cursor

All parameters are optional. Returns: a list of reviews, each with id, title, status, project info, pages (with per-page comment counts), and timestamps.

`list_projects`

List projects in the current org. Returns name, thread count, and review count. Parameters:

Parameter	Type	Default	Description
`limit`	number (1–100)	`50`	Results per page
`cursor`	string	—	Pagination cursor

All parameters are optional. Returns: a list of projects, each with id, name, threadCount, reviewCount, and createdAt.

`search_discussion`

Search chat discussion threads linked to a review. Returns messages from the latest thread by default. Use query to search within a thread and threadId to target a specific linked thread. Returns messages only when the review owner has turned on Include linked chats in Share Settings and the linked thread has messages. Parameters:

Parameter	Type	Default	Description
`reviewId`	string (UUID)	—	Review ID to search discussions for (required)
`query`	string	—	Search query to filter messages
`threadId`	string (UUID)	—	Specific thread ID to search within

All parameters except reviewId are optional. Returns: matching messages from the review’s linked discussion thread.

Action tools

`apply_comment`

Act on a comment. Returns a multimodal response with page screenshots, element metadata, and step-by-step instructions. Control what’s returned with include. Parameters:

Parameter	Type	Default	Description
`id`	string	—	Comment ID from `search_comments` (required)
`include`	`instructions` \| `crop` \| `discussion` \| `all`	`instructions`	What to return (see below)

include options:

instructions — full page screenshot + element metadata + step-by-step implementation instructions
crop — the full page screenshot plus a cropped element zoom for spatial grounding (falls back to full page only when no element region exists)
discussion — the chat discussion behind the comment; requires Include linked chats in the review’s Share Settings, otherwise returns empty
all — instructions, crop, and discussion combined

Publishing tools

`publish_review`

Your agent presents its work as a live web page anyone can open. Reviewers pin comments directly on the page, and each comment they resolve becomes a decision your agent can query and act on. Use it for a prototype your team needs to see or a walkthrough of what the agent built. Supports Vite+React projects, pre-built static sites, and any web project with an index.html. Returns a reviewId (save it — use with search_comments to check for reviewer feedback) and shareUrl (shareable link for reviewers). Parameters:

Parameter	Type	Default	Description
`title`	string	—	Artifact title. Required unless `publishReviewId` is present
`files`	array	`[]`	Source files as `{path, content}` objects (no `node_modules`). Include `package.json` with a `dev` script; files mode builds in the sandbox. For pre-built static sites use `sourceArchive`. Max 500 files, 25 MB total
`assets`	array	—	Binary assets as `{path, filename, contentType, sizeBytes}` objects for two-phase upload (see below). Max 50 assets, 50 MB each
`sourceArchive`	object	—	`{localPath?, sizeBytes?, contentSha256?}` describing a `tar.gz` of the project. Mutually exclusive with `files`, `assets`, and `publishReviewId`. Local `lace-mcp`: set `localPath` and it reads + uploads the file. Hosted MCP: set `sizeBytes` + `contentSha256` to receive a presigned upload URL (see Archive flow). Archives with `package.json` build in the sandbox; pre-built archives with root `index.html` and no `package.json` are served from the CDN. Max 500 MB source, 1 GB static
`publishReviewId`	string (UUID)	—	Review ID from a prior asset-upload prepare phase (see below)
`sourceCommentIds`	array of UUIDs	—	Comment IDs this prototype addresses. Links the artifact back to its source comments
`sourceThreadId`	string (UUID)	—	Thread ID for the conversation that motivated this prototype

Each entry in files has:

Field	Type	Description
`path`	string	File path relative to project root (no leading `/`, no `..`)
`content`	string	File content as text

Each entry in assets has:

Field	Type	Description
`path`	string	Local asset path
`filename`	string	Asset basename used in source rewriting and upload tracking
`contentType`	string	MIME type for the asset. Accepted values: `image/png`, `image/jpeg`, `image/gif`, `image/webp`, `image/svg+xml`, `image/x-icon`, `font/woff2`, `font/ttf`, `font/otf`, `application/javascript`, `text/css`, `text/html`, `application/json`, `application/wasm`, `application/octet-stream`
`sizeBytes`	number	Exact asset size in bytes (max 50 MB)

Text-only flow: pass title and files, including a package.json with a dev script. Returns reviewId and shareUrl. Two-phase asset upload: pass title, files, and assets. The tool returns publishReviewId, CDN URLs, and presigned upload URLs. Upload each asset, rewrite source file references to the CDN URLs, then call publish_review again with publishReviewId and the rewritten files. Do not pass assets and publishReviewId together. If the finalize step returns missing_assets, re-upload the missing files using the fresh URLs and retry. Archive flow: pass title and sourceArchive. Cannot be combined with files, assets, or publishReviewId. With the local lace-mcp, set sourceArchive.localPath and the server reads and uploads the file for you. On the hosted MCP (which cannot read your disk), set sourceArchive.sizeBytes and sourceArchive.contentSha256 (lowercase hex); the tool returns a publishReviewId and a presigned archive_upload.upload_url. Upload the archive there with the returned headers, then call publish_review again with that publishReviewId to finalize.

Config file locations

Tool	Config file
Claude Code	`~/.claude.json`
Claude Desktop	`~/Library/Application Support/Claude/claude_desktop_config.json`
Cursor	`~/.cursor/mcp.json`
Codex CLI	`~/.codex/config.toml`

Troubleshooting

Hosted connector won't connect

Make sure you finished the browser sign-in step. The connector needs a Lace account, and you’ll only see results once you’ve resolved at least one comment.

Local install can't connect to Lace

The Lace desktop app must be open and signed in for the local npx install. Open the app and retry. (The hosted connector does not need the app open.)

I connected the wrong Lace account

In Claude Code, clear Lace authentication from /mcp, then reconnect; Lace will ask you to choose an account again. In clients that do not expose auth reset controls, remove and re-add the Lace MCP connector so the client starts OAuth again. For the local CLI, run npx lace-mcp login --switch.

No comments returned

You need at least one resolved comment. Open Lace, switch to Review mode, pin a comment, and resolve it. Or start from the agent side: ask it to publish something with publish_review, open the share link, comment on the page, and resolve it.

No sessions returned

Start a capture session in the Lace desktop app (header control or ⌘S), work through a few screens, then stop or wait for idle close. Your own sessions appear automatically; a teammate’s sessions appear only if they enabled sharing on that session.

Agent doesn't recognize the MCP server after install

Restart your coding tool fully (quit and reopen), then confirm the Lace server appears in your MCP settings.

Something isn't working — how do I diagnose?

Run npx lace-mcp health from your terminal. It checks connectivity to the Lace backend, validates your credentials, and reports any issues.

Screenshots are missing from comment details

For the local install, screenshots are stored alongside the Lace app data. If they were deleted or the app was reinstalled, the comment will still return card and targeting data but without images.

Privacy and terms

The Lace MCP server reads the comments and decisions your team captures, the recorded sessions you share (or own), and the context behind them. Writes happen only through publish_review, which turns your agent’s work into a shareable review page. It never edits or deletes existing data. See our Privacy Policy and Terms. Questions or support: hello@inlace.co.

Start

Concepts

Agents

Reference

Connect in Claude

Install from the CLI

How it works

Technical reference

MCP server

Session tools

`search_sessions`

`get_session_flow`

Discovery tools

`search_comments`

`list_reviews`

`list_projects`

`search_discussion`

Action tools

`apply_comment`

Publishing tools

`publish_review`

Config file locations

Troubleshooting

Privacy and terms

​Connect in Claude

​Install from the CLI

​How it works

​Technical reference

​MCP server

​Session tools

​search_sessions

​get_session_flow

​Discovery tools

​search_comments

​list_reviews

​list_projects

​search_discussion

​Action tools

​apply_comment

​Publishing tools

​publish_review

​Config file locations

​Troubleshooting

​Privacy and terms

Connect in Claude

Install from the CLI

How it works

Technical reference

MCP server

Session tools

`search_sessions`

`get_session_flow`

Discovery tools

`search_comments`

`list_reviews`

`list_projects`

`search_discussion`

Action tools

`apply_comment`

Publishing tools

`publish_review`

Config file locations

Troubleshooting

Privacy and terms