Skip to main content
Glama

Homespun

Human-in-the-loop for AI agents. Your agent hands a human a real UI — a form, a picker, a dashboard, a diff to review — by URL, and gets the answer back as structured data. No GUI host app, no public address on the agent's side. Works from a cron job, a Slack bot, CI, or any headless server.

Your agent can do anything except ask a human a real question. Homespun fixes that.

npm license stars

Homespun round-trip demo: an agent creates an app, a human approves a release in the browser, and the structured result lands back in the agent's terminal

Works with any agent that can run a shell command or call HTTP — Claude Code & the Claude Agent SDK, LangGraph / CrewAI / OpenAI Agents SDK, or your own cron / CI / bot.

Quickstart

Register once, claim your agent, then deploy:

npx @homespunapps/cli agent register --name "my-agent"   # one-time, hosted relay

A freshly registered agent is unclaimed and cannot deploy yet. Claim it so it is bound to you: open https://homespun.dev/get-started, generate a claim code, then run:

npx @homespunapps/cli agent claim <code>
npx @homespunapps/cli deploy ./my-app   # Node 20+, reads ./my-app/index.html + manifest.json

New apps are private by default: only you and the people you invite can open them. Pass --visibility link or --visibility public to share wider.

Related MCP server: mcp-interactive-ui-server

Reach for Homespun when a text reply is the wrong shape

  • Approvals — deploy gate, refund, PR merge: agent pauses, human clicks approve/reject, agent continues.

  • Forms & pickers — collect structured input instead of parsing prose.

  • Doc / diff review — human marks up a diff; agent gets per-line comments back.

  • Dashboards & status — view-only apps the human just reads.

  • Lists & boards — todo lists, checklists, kanban (records), mutated live.

Try it in 60 seconds ↓ · How it works · Agent reference

The problem

Apps are built for everyone; apps are built for you. Agents can already emit rich output (the "ask Claude for HTML, not Markdown" pattern). But the human's reply is still prose. The agent → human channel is rich, the human → agent channel is a text box. Homespun closes the loop: agent renders a UI (form, picker, doc-review view, dashboard, sketchboard), human manipulates it, every interaction emits structured data, the agent retrieves it (or pushes its own updates back into the same UI). The human "answers" by using a UI, not by typing.

This matters most for agents that live outside a GUI host app: cron agents, Slack/Telegram bots, CI agents, headless servers, personal-agent setups. None of them can use MCP Apps (which needs a host app to render the UI). Homespun needs neither a host app nor a public address on the agent's side; the agent only makes outbound calls to the relay.

Templates and apps — the model

Two nouns carry the whole system:

  • A template is a reusable UI definition: the HTML, an optional event schema, and an optional input schema. Author it once.

  • A app is one use of a template — one context, one (or more) humans, one event log, one TTL. Many apps per template.

The intended flow is author-once, instance-many: register a template with homespun template create, then spin up an app each time you need it with homespun create --template-id <slug> — no HTML re-sent, no regeneration. Per-instance data (the "which PR is this?" data that makes one PR-review template render this PR) rides in --input-data, which the page reads as window.homespun.inputData. For a genuine one-off you can inline the HTML on homespun create --template '<...>' and skip the named template entirely.

A template with no event schema is view-only — a report, dashboard, or chart the human only reads. Give it an event schema when you need an answer back.

How it works

  1. Agent authors (or reuses) a template — an HTML page, optionally with an event schema declaring what the page and the agent may emit.

  2. Agent → POST /v1/apps with {template_id | template, input_data, ttl} → gets {app_id, urls, tokens, expires_at}.

  3. Agent delivers urls.humans[0] to the human over whatever channel it already has.

  4. Human opens the URL. The relay serves a small shell page that loads the template in a sandboxed iframe (locked-down CSP), plus a tiny app runtime exposing window.homespun — homespun.emit(type, data), homespun.on(type, handler), homespun.state, homespun.records, homespun.inputData.

  5. Human interacts → each homespun.emit(...) is POSTed to /v1/apps/{id}/events → appended to that app's event log (validated against the schema; a wrong shape or wrong author is rejected).

  6. Agent retrieves: stream over the WebSocket (homespun watch), long-poll (GET /v1/apps/{id}/events?since=<cursor>&wait=<s> / homespun show --wait), or register a webhook. The "ask the human" call blocks until the awaited event or a timeout.

  7. Homespun expires after ttl.

Event ordering. A client connected over the WebSocket may receive an event via the broadcast stream before the ack for its own write of that same event. Clients de-duplicate on the event id — treat the id, not arrival order, as the source of truth.

More than the round trip

The core round trip is the foundation; the relay also carries the pieces you need to build real, durable agent↔human surfaces:

  • Reusable templates + a marketplace. homespun template create / list / search / show / version manage named, versioned templates. A human can mark one public in the relay's web UI, giving it a listing other humans can browse and install into their own account.

  • Records — per-app mutable collections (posts, comments, kanban cards, line items) keyed by stable record_key, with optimistic locking and soft-delete. Use records when the current value matters and history doesn't; use events when history is the point. Declared with a record_schema (JSON Schema 2020-12 + the x-homespun-collections extension).

  • Attachments — upload images/PDFs/audio/video (homespun attachment …), reference them by id in events, and let the page fetch them lazily or accept uploads back from the human. Capability-URL (/b/<token>) downloads, MIME sniffing, and EXIF stripping included.

  • SQL query — homespun query "<SQL>" runs read-only DuckDB SQL (a PostgreSQL-compatible dialect) scoped to your own apps, records, and events.

  • Multi-participant apps — beyond the single auto-minted human URL, an owner can add identity-bound (email-invite) or public/anonymous participants. homespun participant list / new / revoke manage URLs on a live app.

  • Taste & feedback — homespun taste remembers a human's presentation preferences across runs so authored UIs stay consistent; homespun feedback reports issues about app itself.

See skills/homespun/SKILL.md — the agent-facing reference — for the authoritative, version-matched description of all of the above.

How Homespun compares

Homespun is not a competitor to in-chat UI extensions — it owns a different quadrant. MCP Apps (SEP-1865), MCP elicitation, and AG-UI / CopilotKit all render UI inside a live host or chat session: a host app (Claude Desktop, ChatGPT, VS Code) or your own front-end is present, the human is looking at it, and the UI is torn down when the turn ends. Homespun is for the agents those approaches can't serve — the ones with no session at all: a cron job, a CI pipeline, a Slack or Telegram bot, a headless server, or any case where the human is on a different device than the agent. Agents that live outside a GUI need a callback URL, not a chat window.

Homespun

MCP Apps (SEP-1865)

MCP elicitation

AG-UI / CopilotKit

Where the UI renders

Standalone URL, any browser/device

Sandboxed iframe inside the MCP host app

Inside the MCP client (form/URL prompt)

Inside your front-end app

Needs a host app / live session

No — agent makes outbound HTTP only

Yes — an MCP host renders it

Yes — an MCP client must be connected

Yes — a running front-end + SSE session

Out-of-band delivery (URL to another device/channel)

Yes — hand the URL over Slack, Telegram, email, SMS

No — bound to the host UI

No — the client drives the prompt in-session

No — bound to the app the user is in

Survives the agent turn ending

Yes — app lives until its TTL; the human can answer later

No — tied to the conversation turn

No — request is resolved within the call

No — tied to the live session

Persistent / mutable state

Yes — durable event log + mutable records per app

Per-render component state

Single request/response

Session state (STATE_DELTA), app-managed

Self-hostable

Yes — one container, MIT

Spec only; depends on host

Spec only; depends on client

Open-source; you host the app + agent

The rows app wins on are exactly the headless/out-of-band ones: no host app, URL-on-another-device delivery, an answer that can arrive minutes or hours later, and state that outlives a single turn. If you do have a live host or front-end session, MCP Apps / elicitation / AG-UI are the natural fit — render the UI right where the human already is. Homespun is the tool for everywhere else.

Examples

Runnable, copy-pasteable examples live in examples/:

  • claude-code-approval/ — a CLI agent (Claude Code or any shell agent) hands a human an "approve this plan?" app and reads the decision back from homespun watch.

  • ci-deploy-gate/ — a GitHub Actions deploy gate: the pipeline posts an app URL, a human approves/rejects with a reason, and a script polls the result and exits 0/1.

Install

No build step, no host app. Pick your audience — paste the agent block into your AI agent's chat, or run the human block yourself.

Paste to your AI agent

Paste the block below into your AI agent's chat. It will install the CLI, register against the hosted relay, and install the Homespun skill into its own skill directory.

Install Homespun for me. Homespun lets you (the agent) build me (the human) an app: a rich interactive UI hosted by URL, built for the moment I need it. If it captures input, you get my response back as structured data. Use it whenever a text reply is the wrong shape: forms, pickers, dashboards, doc or diff review, approvals, status pages, multi-step wizards, file uploads — anything better seen than described. Input is optional; a read-only dashboard or status view is a valid Homespun.

Install everything at GLOBAL / USER scope, not in the current workspace. Homespun is a personal capability I want available in every project, not a per-repo dependency. Do not create files in the current working directory or commit anything to this repo.

Do the following:
1. Install the CLI globally: `npm i -g @homespunapps/cli` (needs Node 20+). The binary is `homespun`. Do not add it to the current project's package.json.
2. Install the Homespun skill into your agent's GLOBAL / USER skill directory, not a project-local one. For example:
   - Claude Code: ~/.claude/skills/homespun/SKILL.md  (NOT .claude/skills/ in the current repo)
   - Cursor: user-level rules, NOT .cursor/rules/ in the current repo
   - Codex / Gemini / Copilot: your user-scope AGENTS.md / GEMINI.md / instructions file, NOT the project root
   Run `homespun skill show` to fetch the skill from the relay you're targeting — it prints the raw markdown to stdout, version-matched to that relay. Capture that output and save it in the appropriate format for your runtime (Claude Code keeps the markdown as-is; Cursor wants `.mdc` rules; Codex / Gemini compose it into the user-scope AGENTS.md / GEMINI.md).
3. Register against the hosted relay: `homespun agent register --name "<short-descriptive-agent-name>"`. Pick a name that identifies your agent (e.g. `claude-code-lalit-macbook`, `cursor-work-laptop`) — the default is unhelpful. One-time, writes a key to ~/.config/homespun/config.json (already user-scope).
4. Confirm install with `homespun key list`. This round-trips your key against the relay and prints { agent_id, name, key_prefix, ... } — proves the CLI is installed AND registration succeeded.
5. Now exercise the round-trip end-to-end so I can see it work:
   - Create a simple two-way app — a small inline HTML template with a text input and a submit button that emits an event back to you with `homespun.emit("hello", {msg})` (use whatever shape `homespun create --help` and the SKILL.md describe; the simplest "hello" form is fine).
   - Print the resulting app URL (`urls.humans[0]`) and ask me to open it.
   - Start watching for the submit event (e.g. `homespun watch <homespun-id> --type hello`).
   - When I submit, show me the structured event payload you received. That's the "aha" moment — confirms the full round-trip works.

After this, reach for `homespun create / show / send / watch` (and `homespun template …` to reuse UIs) whenever a UI would communicate better than text. Run `homespun <command> --help` for authoritative options.

Run yourself (human)

Six commands. Needs Node 20+.

# 1. Install the CLI (Node 20+)
npm i -g @homespunapps/cli

# 2. Register with the hosted relay, pick a short, descriptive agent name
homespun agent register --name "<short-descriptive-agent-name>"

# 3. Confirm, round-trips your key against the relay
homespun key list

# 4. Install the skill into your agent. `homespun skill show` prints the relay's
#    current, version-matched skill markdown to stdout, save it where your
#    runtime keeps skills (e.g. ~/.claude/skills/homespun/SKILL.md for Claude Code).
homespun skill show > ~/.claude/skills/homespun/SKILL.md

# 5. Claim your agent so it can deploy. Open https://homespun.dev/get-started,
#    generate a claim code, then run:
homespun agent claim <code>

# 6. Deploy your first app (reads ./my-app/index.html + manifest.json).
#    New apps are private by default: only you and the people you invite can
#    open them. Pass --visibility link or --visibility public to share wider.
homespun deploy ./my-app

Use from an AI chat (remote connector — no install)

No terminal? Add Homespun to Claude (web, desktop, mobile), ChatGPT, or any chat app that supports remote MCP connectors — nothing to install. Add a custom connector pointing at:

https://homespun.dev/mcp

Log in with your email (magic link), approve the consent screen once, and Homespun's tools show up in the chat — then ask it to "build me an app for …" and it hands back a URL. This is the only way to drive Homespun from a phone or a pure chat app.

Full walkthrough (Claude web/desktop/mobile, Claude Code's claude mcp add, ChatGPT, and others) → docs/CONNECT-AI-CHAT.md.

Distribution

The repo is an npm-workspaces monorepo with four packages:

  • @homespunapps/core — the relay client: a pure, framework-free HTTP + WebSocket library (HomespunClient + openStream). Build any client on it.

  • @homespunapps/relay — the relay server. Use the hosted instance, or self-host it with docker compose up (bundled Postgres) — see Self-hosting.

  • @homespunapps/cli — the homespun command-line tool. The agent's entry point: emits JSON on stdout, so it's harness-agnostic — works for an MCP host, a cron agent, a shell pipeline, a CI job, or a process supervisor. homespun watch <id> --type <event> streams an app as JSON-lines and exits when the awaited event lands. A LangChain tool wrapper may come later.

  • @homespunapps/mcp — a thin stdio MCP server (binary homespun-mcp) with full parity with the CLI. Point Claude Desktop, Cursor, or any MCP client at npx @homespunapps/mcp and Homespun shows up as tools: the hot-path discrete tools (create_app, get_events, send_to_app, update_app, upgrade_app, list_apps, delete_app, the record CRUD tools) plus consolidated action-based tools for templates, sharing, attachments, taste, keys, trash, feedback, agent identity, and run_query / get_skill. See the package README for the full tool list and client config snippets.

The CLI is homespun <command> [options] — create, show, send, watch, list, delete, and participant operate on an app; template, attachment, records, query, taste, feedback, key, agent, config, and skill are the other command groups. Run homespun --help for the full list.

Use from an MCP client (local stdio server)

Run Homespun as a local stdio MCP server inside any MCP host (Claude Desktop, Cursor, …) — no global install needed. (For a chat app with no terminal, or your phone, use the remote connector instead — nothing to run locally.)

{
  "mcpServers": {
    "homespun": {
      "command": "npx",
      "args": ["-y", "@homespunapps/mcp"],
      "env": { "HOMESPUN_API_KEY": "hs_..." }
    }
  }
}

Omit env to let the server auto-register an agent on first use. Full setup, the tool list, and the poll-for-events pattern are in the @homespunapps/mcp README.

Stack

TypeScript. Runtime: Node 20+ (Bun fine too). Web: Hono (tiny, fast, container/edge-friendly). ORM: Prisma. PostgreSQL, self-host and hosted alike. npm workspaces for the monorepo. See docs/SPEC.md.

Self-hosting

You don't have to run a relay — point the CLI at the hosted instance and you're done. But Homespun is open-core (MIT) and self-hosts with no paid dependencies:

  • docs/SELF-HOSTING.md — run your own relay with docker compose up (relay + bundled Postgres). Set three env vars, done.

  • docs/DEPLOY.md — the operator guide: managed Postgres, multi-replica scaling, observability, and the Azure Container Apps reference deployment.

The relay is configured entirely through environment variables — packages/relay/.env.example is the full reference.

Contributing

Issues, fixes, and design feedback are welcome. See CONTRIBUTING.md for dev setup, the test suites, and PR conventions, and CODE_OF_CONDUCT.md for community expectations. Security vulnerabilities: please report them privately — see SECURITY.md.

See also

  • skills/homespun/SKILL.md: the agent-facing reference — every command, the template/app model, schemas, records, attachments, query (authoritative, version-matched to the relay)

  • docs/CONNECT-AI-CHAT.md: add Homespun to Claude / ChatGPT / any chat app as a remote connector — no install

  • docs/SPEC.md: technical design (architecture, API, data model, bridge, auth, open/closed split)

  • docs/SELF-HOSTING.md: run your own relay with docker compose up (relay + bundled Postgres)

  • docs/DEPLOY.md: operator deployment — managed Postgres, scaling, observability, Azure

  • docs/ROADMAP.md: scope, later phases, strategy notes

  • docs/architecture/: per-phase implementation docs (Prisma models, endpoints, the app runtime, the CLI)

  • Prior art / landscape: MCP Apps (blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps/), mcp-ui (github.com/MCP-UI-Org/mcp-ui), AG-UI (copilotkit.ai), A2UI (Google), Thesys C1

  • Motivating read: Thariq, "Using Claude Code: The Unreasonable Effectiveness of HTML" (simonwillison.net/2026/May/8/unreasonable-effectiveness-of-html/)

Available Tools

26 tools
agentManage Agent IdentityA
Destructive

Agent identity + binding. ONE tool with an action enum: whoami (the resolved relay URL, active profile, whether a key is configured — no network, no secrets) | claim (bind this agent to a human via a one-shot claim code from their Settings UI; one-way) | logout (clear the locally-saved key/profile; does NOT revoke it on the relay — use the key tool's revoke for that).

Inputs: action (required), code.

attachmentsManage AttachmentsA
Destructive

Binary attachments (images, PDFs, audio, video) referenced from event payloads / input_data via format: pane-attachment-id. ONE tool with an action enum: upload | download | show | list | delete | mint_token | revoke_token | list_tokens. upload reads an ABSOLUTE file_path; download writes to an ABSOLUTE out_path (or returns base64). Scope an upload to agent (default, reusable), pane, or template. mint_token returns a /b/ capability URL (ONCE) a browser can GET without your API key.

Inputs: action (required), mime, once, limit, +10 more.

create_paneCreate PaneA
Destructive

Hand the human a rich interactive UI by URL and (optionally) get structured data back. Build the UI as inline HTML (pass name + html) OR reuse a saved template (pass template_id). The relay hosts it and returns a URL. ALWAYS give the returned url to the human — paste it into the conversation and ask them to open it. Reach for this whenever a text reply is the wrong shape: forms, approvals, pickers, surveys, dashboards, diff/doc review, wizards. If the page captures input it emits events back to you (poll them with get_events) or mutates record collections (the record tools). BEFORE authoring: call get_skill for the events-vs-records decision + schema grammar, and the taste tool (action: get) for the human's house style — both shape the HTML you write. Returns { pane_id, url, urls, title, expires_at }.

Inputs: html, name, tags, title, +14 more.

delete_paneDelete PaneA
DestructiveIdempotent

Close/delete a pane (idempotent — an already-closed pane still succeeds). The human's URL stops working. To merely edit a pane keep it alive with update_pane; to recover a soft-deleted pane use the trash tool (action: restore).

Inputs: pane_id (required).

delete_recordDelete RecordA
DestructiveIdempotent

Soft-delete a row from a pane's record collection. The page sees the deletion live (the row becomes a tombstone in list_records). Pass if_match for an optimistic-locked delete. Returns { deleted: true }.

Inputs: pane_id (required), collection (required), record_key (required), if_match.

delete_record_collectionDelete Record CollectionA
DestructiveIdempotent

Drop a WHOLE per-pane record collection at once: every row plus the collection row itself. Use this to reset or remove a collection (todo list, comment thread, board) rather than deleting rows one by one with delete_record. Owner-only and destructive, so it requires confirm:true. Collection names are immutable, so to rename a collection drop the old one and write under the new name. Returns { deleted: true, collection }.

Inputs: confirm (required), pane_id (required), collection (required).

feedbackManage FeedbackA
Destructive

Send or list feedback to the relay operator. ONE tool with an action enum: create (a bug|feature|note with a message, optional pane_id) | list (the agent's own submissions, newest first, paginated by before).

Inputs: action (required), type, limit, before, +2 more.

get_eventsGet EventsA
Read-only

Poll a pane's append-only event log for what the human did (form submissions, approvals, picks). This is how you receive the round-trip result — there is no push/streaming in MCP. Poll loop: call with no since first; process the returned events; remember next_cursor; call again passing it as since to get only newer events. To WAIT for a human who hasn't acted yet, pass wait_seconds (~25) so the relay holds the request open until an event arrives or it times out, then call again with the same cursor. Returns { events, next_cursor }.

Inputs: pane_id (required), since, wait_seconds.

get_pane_stateGet Pane StateA
Read-only

Fetch a pane's current metadata (status, title, template version, timestamps, expires_at) WITHOUT its event log. Use it to check whether a pane is still open or has expired. To read what the human did, use get_events.

Inputs: pane_id (required).

get_recordGet RecordA
Read-only

Fetch a single record row by its key from a pane collection (scans the collection — fine for a one-off lookup, not a hot loop). Returns { record } or an isError record_not_found.

Inputs: pane_id (required), collection (required), record_key (required).

get_skillGet Skill GuideA
Read-only

Fetch the relay's auto-updating SKILL.md (the full Pane usage guide) — UNAUTHENTICATED, needs no API key. Call this to self-teach the Pane workflow (events vs records, schema grammars, the poll loop) before driving the other tools. Pass version_only:true to get just the relay's skill version string (to check if a cached copy is stale).

Inputs: version_only.

keyManage API KeyA
Destructive

Inspect or revoke the calling agent's API key. ONE tool with an action enum: list (key info — agent_id, key_prefix, timestamps) | revoke (self-destruct the agent's OWN key; it stops working immediately and is irreversible — pass confirm:true). The relay scopes keys to the caller, so both act only on your own key.

Inputs: action (required), confirm.

list_panesList PanesA
Read-only

Enumerate YOUR agent's panes (newest first). Use it to find a pane_id you lost, audit what's open, or get a cursor for pagination. No secrets in the response (participant tokens are unrecoverable — mint a fresh URL with the participant tool). Filter by status (open|closed|all) or template_id. Returns { items, next_cursor }.

Inputs: limit, cursor, status, template_id.

list_recordsList RecordsA
Read-only

List rows in a pane's mutable record collection (todo list, shopping list, kanban board, comment thread). Records are the right primitive when the page shows several mutable items and the CURRENT state matters more than the history. This also doubles as the POLL/watch for records (no streaming in MCP): pass the prior next_since to fetch only newer/changed rows. include_tombstones:true surfaces deletions. Returns { records, next_since, has_more }.

Inputs: pane_id (required), collection (required), limit, since, +1 more.

participantManage ParticipantsA
Destructive

Manage a pane's participant URLs (recovery + leak-containment). ONE tool with an action enum: list | new | revoke. Use new when you lost the original URL (the plaintext token is returned ONCE — save it). Token URLs are stored hashed and cannot be recovered.

Inputs: action (required), pane_id (required), participant_id.

run_queryRun SQL QueryA
Read-only

Run read-only SQL over YOUR scoped data (panes, records, events) — the relay scopes every row to panes you own. Use it to summarise activity, find panes/records by content, or build a report. Tables + columns and JSON projection operators are documented on the sql parameter. Default output is { columns, rows, truncated, scope, elapsed_ms } (format:json); csv/tsv/table render the rows as text. Capped at 10,000 rows; 10s timeout.

Inputs: sql (required), format, pane_id.

send_to_paneSend to PaneA
Destructive

Push an event INTO an open pane — update the live UI the human is looking at (progress, a new message, a status change, fresh data). The event type must be declared in the pane's event_schema with 'agent' in its emittedBy. For mutable collections (todos, line items, comment threads) prefer the record tools instead. Returns { event, deduped }.

Inputs: data (required), type (required), pane_id (required), idempotency_key.

shareManage Pane SharingA
Destructive

Identity sharing on a pane (layered on top of participant tokens). ONE tool with an action enum: list (access_mode + grants) | invite (a human by email, role participant|viewer) | set_access (the /p access mode: invite_only|link|public) | revoke (one grant by id). Token (/s/) links are independent of access_mode and keep working.

Inputs: action (required), pane_id (required), role, email, +2 more.

tasteManage UI Taste NotesA
Destructive

Read / write / clear the agent's freeform UI taste notes (a small markdown document of presentation preferences learned from human feedback — 'denser layout', 'no rounded corners'). ONE tool with an action enum: get | set | clear. Call get BEFORE generating a pane so prior feedback shapes the output; set does a whole-document replace (not append). Keep entries about UI/presentation only.

Inputs: action (required), taste.

templateManage TemplatesA
Destructive

Manage reusable, versioned UI templates (author once, instance many times via create_pane's template_id). ONE tool with an action enum: create | version | update | search | list | show | get_version | delete | publish | unpublish | search_public | set_icon. Required fields per action are documented on the action parameter. A template is HTML + an event schema (+ optional input/record/template-record schemas); a pane is one use of one version of it.

Inputs: action (required), id, html, name, +17 more.

template_recordsManage Template RecordsA
Destructive

CRUD for TEMPLATE-level record collections — owner-curated content anchored to a template head and visible to every pane derived from any of its versions (vs per-pane records, which are the discrete record tools). ONE tool with an action enum: list | get | upsert | update | delete | delete_collection. The template version must declare the collection via template_record_schema (set it with the template tool first).

Inputs: action (required), collection (required), template_id (required), data, +6 more.

trashManage TrashA
Destructive

Manage soft-deleted panes + templates. ONE tool with an action enum: list | restore (pane id) | restore_template (template id|slug) | purge (pane id) | purge_template (template id|slug). purge bypasses the retention window and is permanent. Soft-deleted rows live in trash until the sweeper reclaims them.

Inputs: action (required), id.

update_paneUpdate PaneA
DestructiveIdempotent

Edit instance-level fields on a LIVE pane in place (PATCH) without minting a new one — the pane keeps its id, URL, event log, and template pin. Settable: ttl_seconds OR expires_at (mutually exclusive), title, preamble, input_data (replaced wholesale + revalidated), metadata, tags, icon_emoji / icon_attachment_id (or clear_* to drop the override). Pass at least one field. Returns the full new pane state + an updated_fields array. To swap the HTML/schemas, use upgrade_pane instead.

Inputs: pane_id (required), tags, title, metadata, +8 more.

update_recordUpdate RecordA
DestructiveIdempotent

Update an existing row in a pane's record collection (replaces its data). Pass if_match with the row's current version for an optimistic-locked update — on a version mismatch the relay returns the current row so you can retry. Returns { record }.

Inputs: data (required), pane_id (required), collection (required), record_key (required), +1 more.

upgrade_paneUpgrade PaneA
Destructive

Re-pin a LIVE pane to swap its HTML (design) + event/input/record schemas in place — same URL, no new pane. Two ways: (1) pass html to EDIT AN INLINE PANE'S HTML in one call — the relay appends a fresh version with that HTML and re-pins (schemas you omit are inherited from the current version, so to change only the HTML pass only html); inline panes only. (2) pass template_version to re-pin to a version you already appended with the template tool (action: version) — for named/reusable templates. By default a strict schema-compat gate refuses an upgrade that would narrow the schema (returns schema_incompatible_upgrade + details.breaks); pass force:true to apply anyway. Returns { pane_id, template_version, upgraded, breaks, compat }.

Inputs: pane_id (required), html, force, event_schema, +5 more.

upsert_recordUpsert RecordA
DestructiveIdempotent

Create a row in a pane's record collection, or return the existing row if record_key is already present (deduped:true). Use to add a todo, a line item, a comment, etc. The collection must be declared in the pane's record schema with 'agent' allowed to write. If you're still designing the pane, call get_skill first for the records-vs-events decision and the x-pane-collections schema grammar. Returns { record, deduped }.

Inputs: data (required), pane_id (required), collection (required), record_key.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 26 tools

    Disambiguation4/5

    Tools cover distinct aspects of the Pane system: identity, attachments, pane lifecycle, records, templates, etc. Some overlap exists between update_record and upsert_record, and between delete_pane and trash, but descriptions clarify the differences. Overall, an agent can reasonably distinguish between tools.

    Naming Consistency3/5

    Mixed naming conventions: some tools use verb_noun (create_pane, delete_pane, list_records), while others are single-word nouns (agent, attachments, feedback, key, trash, taste) with action enums inside. This inconsistency could cause confusion, though each tool's purpose is clear.

    Tool Count3/5

    26 tools is on the higher side but not unreasonable for the server's scope covering panes, records, templates, attachments, identity, and more. Some tools could potentially be consolidated (e.g., trash could be part of pane/template operations), but the count is still manageable.

    Completeness4/5

    The tool surface covers CRUD for panes, records, templates, and attachments, plus identity management, event polling, and feedback. Minor gaps exist (e.g., no explicit tool to list all templates requires using template's search action), but core workflows are well-supported.

    Maintenance

    ActivityNo data
    ResponsivenessSyncing

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI clients to create multi-field forms, send approval requests via email, and retrieve structured human responses. It facilitates human-in-the-loop workflows by allowing agents to collect data and check form completion status.
      2,530 npm
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      Enables AI agents to render interactive user interfaces such as forms, dashboards, charts, tables, and wizards directly in MCP-compatible clients. Supports structured data collection and richer interactions beyond text responses.
      6
      13 PyPI
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to submit tasks for human or AI review and receive decisions via MCP tools, adding human review checkpoints to workflows.
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to ask clarification questions and receive structured user input through a Human-in-the-Loop interface.
      1
      MIT