collection_create
ChatGPTCreate a named collection in a workspace, to file projects under (project_create or project_start_session with collectionId, or project_move).
collection_get
ChatGPTA collection by id, as its share link shows it: its projects, its skills and its files. Use it when the user hands you a collection link.
collection_remix
ChatGPTCopy a whole collection — every project, its skills and its files — into a workspace (orgId). Tell the user how many projects it copies and how much room their plan has, then call again with confirm: true.
component_delete
ChatGPTDelete one component from a project (it stays restorable for a while from the workbench). The last one cannot be deleted. Requires ifRevision.
component_save
ChatGPTSave one component's metadata: name, story (its controls), or values (a merge into the story's values — what the controls are set to). Without componentId it creates a new component in the project. File contents go through a code session, never here. Requires ifRevision.
file_get
ChatGPTRead one file of a tree (collectionId, or null for the workspace root) by path: its mime and size, a signed url of its exact bytes (curl it; it expires in an hour), and content inline for a text file up to 256 KiB. To use the file in a fragment, copy it into the fragment — fragments never reference collection files. Follow next.
file_save
ChatGPTAdd one file to a tree (collectionId, or null for the workspace root) at path — folders come from the path ("logos/mark.svg"). A file already at that path (ignoring case) is replaced; there is no history. Pass content (UTF-8 text) or contentBase64 (bytes), up to 16 MiB. Pass neither to get a signed upload url and a curl command instead — use it for anything large or already on disk (up to 25 MiB). Any member may write; counts toward the plan's storage.
files_find
ChatGPTFind files across a workspace's file trees (every collection's and the root's; archived collections left out) by name or path (query, any case) and type (mime: exact, or a family like image/*). Answers each hit with its collectionId (null: the root), path, mime and size — read one with file_get. Not a fragment search: that is fragment_find. Reads only.
files_list
ChatGPTList one file tree — a collection's (collectionId) or the workspace root's (collectionId: null): every file's path, mime and size, and the folders (empty ones too). prefix narrows to one folder. Files are context for you — logos, PDFs, fonts, reference data; read one with file_get, or write the whole tree to disk with files_pull. Reads only.
files_pull
ChatGPTWrite a whole file tree (or one folder of it: prefix) to disk: answers every file with a signed url and one script.command (curl … | sh -s -- <dir>) that creates the folders and downloads every file under <dir>. Use it when the project needs the collection's files — brand assets, fonts — on disk. URLs expire in an hour. Follow next.
fragment_app_activity
ChatGPTThe preview card's read of who is working on the fragment (GET /fragments/:id/activity). Called by the card, not the model.
fragment_app_get
ChatGPTThe preview card's read of the fragment (GET /fragments/:id). Called by the card, not the model; use fragment_get.
fragment_app_preview
ChatGPTThe preview card's live-preview grant: SHORT-lived (an hour; the card re-asks ahead of it). Called by the card, not the model; use fragment_preview.
fragment_app_save_values
ChatGPTThe preview card's Save: merges top-level story values into a variant (null deletes). Called by the card, not the model; use fragment_save_variant. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_archive
ChatGPTArchive a fragment, removing it from workspace lists and taking its public page and share link offline. Asks the user first: without confirm: true it answers confirmation_required. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_collection_create
ChatGPTCreate a collection — a named group fragments are filed under (one level; a collection holds fragments, never other collections). Answers its id; pass it as collectionId to fragment_start_session / fragment_create, or file existing fragments with fragment_move. Check fragment_list first: reuse a collection that already fits rather than creating a near-duplicate.
fragment_collection_get
ChatGPTRead a collection (a kit): the collection, its fragments, its skills in full, what you may do (viewer.member, viewer.canRemix) and, per workspace you could remix into, how many fragments it has against the room left there (room). Use it to read a kit from a share link (/share/collection/<id> — the id is the collectionId) as well as your own collections. Read the skills before using the fragments.
fragment_collection_remix
ChatGPTCopy a whole collection into a workspace you belong to: a NEW collection with every fragment (each a remix) and the collection's skills; the source workspace's own skills do not travel. All or nothing against the target's plan cap. Asks the user first: without confirm: true it answers confirmation_required with the fragment count and the room in each workspace — show those numbers, then retry with confirm: true. orgId picks the target (defaults to the source workspace for a member); name defaults to the original plus " (remix)".
fragment_create
ChatGPTCreate an empty fragment (starter variant) without opening a session; answers the fragment and its shellUrl. fragment_start_session without fragmentId is the normal path.
fragment_delete_variant
ChatGPTDelete a variant and its entry file (the last variant is refused: last_variant). The default repoints to a survivor. The variant and its history are kept 30 days: the user can restore it from Version history. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_describe
ChatGPTSet what a fragment is for: description (one line, '' clears) and tags (a few lowercase words; the list replaces the current one, [] clears). Both show on fragment_list and the workspace, so keep them current when the purpose changes. Answers the fragment with its stored (normalized) tags. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_docs
ChatGPTForge Fragments reference, returned as markdown: the reference for cloud-hosted Fragments — using one in a project (fragment_find, the skills to read, fragment_pull, adapting), remixing fragments and collections, skills (SKILL.md documents a workspace or collection carries), the Vite-shaped tree (variants, shared files, assets, package.json, index.css), code sessions (download, edit, upload, submit), limits and path rules, the story body, revisions and restore, publishing, error codes. The server instructions carry the must-know; read this for limits, path rules, revisions and error codes.
fragment_feedback_list
ChatGPTThe user's feedback queue on the fragment's preview: comments pinned to a spot, proposed style edits (CSS changes on an element), screenshots with a note — each with a reply thread and open/resolved status. Check it when you start work on a fragment and after you submit; address open items, answer in the thread (fragment_feedback_reply) and close what you handled (fragment_feedback_resolve). Anchors are fractions of the frame (x, y) at frameWidth plus, when known, the element under the click: selector, tag, text, classes (grep these), components (the fragment's React component chain, outermost first), attributes, computed styles and the surrounding text — enough to find it in source without rendering. Screenshots carry image.url (curl it).
fragment_feedback_reply
ChatGPTAnswer in a feedback thread as the agent (what you did, what you need, why not). Keep it short; resolve separately once the item is handled.
fragment_feedback_resolve
ChatGPTMark a feedback item resolved once your submit addressed it (or reopen: true to put it back). The shell drops resolved pins from the preview.
fragment_find
ChatGPTCall this BEFORE building any piece of UI: it ranks the fragments across your workspaces that fit what you are about to build and names the skills to read first. Pass intent (what you are building, in your words) and, when you know them, project facts (framework, styling, dependencies). Answers ranked (best first, with probability, fits and the governing skillIds), skills (governs: attached to a ranked hit; relevant: judged to apply on its own), a gate (fits: false means build from scratch, still following the relevant skills), ranker (jev, or flat when the plane ranked lexically — then verify the hit yourself) and next, the step to take. Reads only.
fragment_get
ChatGPTEverything about one fragment except file contents: summary and revision, variants (meta, story, entryPath), the projected tree (path, kind, bytes, sha256), package.json dependencies, public state. Read before editing.
fragment_get_file
ChatGPTOne file of the tree. Text files (variants, src/**, index.css, package.json) come back inline as content; binary assets answer downloadUrl (authenticated) and deliveryPath (what the preview serves) instead. Read a file before you rewrite it.
fragment_list
ChatGPTList the fragments in an organization (name, description, tags, ids, revision, public state, sizes, collectionId) and its collections — the named groups fragments are filed under (id, name, fragmentCount). Scan description and tags to find the one the user means. orgId is only needed when you belong to several.
fragment_list_revisions
ChatGPTHistory by fragment revision, newest first: which tree paths each revision changed. With path, only that file. Bounded by per-path retention.
fragment_move
ChatGPTFile a fragment under a collection (collectionId from fragment_list or fragment_collection_create), or pass collectionId: null to return it to the workspace root. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_preview
ChatGPTWhen the user wants to see the fragment: answers shellUrl — the workbench page (live preview, controls, agent status) — which is the URL to hand them; variant (id or name) opens that variant. Also answers url, the raw sandboxed render (valid 12 h, re-renders on every submit) for your own checks such as a screenshot or fetch — never give the user that one — plus the fragment's name and variants. Clients that render MCP Apps (ChatGPT) show the result as a live preview card with a variant switcher, so there you need not repeat the links.
fragment_publish
ChatGPTPublish the current revision at https://<slug>.ui8.dev/ (a pinned snapshot; later edits stay private until the next publish). slug claims or changes the public name. password gates the page behind one shared password (null removes it; the page is open by default). Asks the user first: without confirm: true it answers confirmation_required. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_pull
ChatGPTTake a fragment into the project you are working in: answers its text files inline (or one tar url + command above the inline cap), a signed url per binary asset (curl them; they expire in an hour), the variant's entryPath and story, package.json dependencies, the skills that govern it (most specific first) and provenance.header to put at the top of the adapted entry file. Adapt the code to the project's conventions — nothing is copied in the cloud and the fragment is not changed. Needs membership of the workspace or a remix-level share link. Follow next.
fragment_remix
ChatGPTCopy a fragment as a NEW private fragment (every variant, file and asset; fresh ids, no history, unpublished) and answer it with its shellUrl. Works on a fragment in your workspace, or one reached through a share link whose linkAccess is remix. orgId picks the workspace the copy lands in (defaults to the source workspace when you are a member; required for a share-link remix). name defaults to the original plus " remix". variantId copies just that variant into its own fragment (Extract: the variant stays in the original; the copy is named "<fragment> · <variant>" by default). collectionId files the copy under a collection of the target workspace (null: its root). The copy records where it came from (remixedFrom).
fragment_rename
ChatGPTRename a fragment. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_restore_revision
ChatGPTRestore every variant and file (or one path) to its state at a revision, written as a NEW revision; answers what was restored and warnings for paths whose history does not reach back that far. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_save_variant
ChatGPTVariant metadata — never code. With variantId: name renames, story replaces the whole story body, values merges top-level props (null deletes). Without variantId: creates a variant (name required) with a starter entry file and answers its entryPath — fill it through a session. Returns story warnings to fix. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_set_default_variant
ChatGPTChoose which variant the preview and the public page open by default. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_set_password
ChatGPTPassword-protect the public page (one shared password visitors type once), or pass null to make it open again. Takes effect on a live page at once — no re-publish — and is remembered for the next publish. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_start_session
ChatGPTCall this FIRST for any confirmed create or edit, before research or drafting. Without fragmentId it creates the fragment (name required; pass a one-line description and a few tags too, so the workspace list stays searchable) and its first variant. Answers shellUrl (the fragment's workbench page: live preview, parametric controls and the agent-activity indicator — hand it to the user right away), workdir (a local directory the commands use), the file list (kind, bytes, sha256), download.command (a gzipped tar of the Vite-shaped project, extracted into workdir), upload.command and a deadline (60 minutes). File contents are never inlined: run download.command, read only the files you need, edit under workdir, run upload.command exactly as returned, then fragment_submit with fragmentId and sessionId. For a new fragment the download is optional — write the variant entry file into workdir and upload. Never start a session for questions or read-only work.
fragment_submit
ChatGPTArgs: fragmentId, sessionId, optional deleted (tree paths to remove) and variant (id or name the URLs open). Apply the tree you uploaded to the session (run upload.command first). Diffs the upload against the fragment by content hash and writes only what changed, in one revision; answers at once with the revision, shellUrl (the workbench page — the ONLY URL to give the user), previewUrl (the raw sandboxed render, for your own checks; never hand it to the user), what was written, and ignoredPaths (server-owned and tooling files — never an error). List removals in deleted. A rejected submit keeps the session: fix, re-upload, submit again. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_unarchive
ChatGPTRestore an archived fragment and its public page and share link. Consumes a fragment slot; quota_exceeded leaves it archived. No confirmation needed. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
fragment_unpublish
ChatGPTTake the public page down; the slug is kept so a later publish restores the URL. Asks the user first (confirm: true). Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
list_orgs
ChatGPTThe organizations the caller belongs to. Only needed when a tool answers org_choice_required.
project_archive
ChatGPTArchive a project: it leaves the workspace lists, its public page and share link stop answering, and it stops counting against the plan. Tell the user, then call again with confirm: true. Requires ifRevision.
project_create
ChatGPTCreate an empty project (a starter cover component) without opening a session. Prefer project_start_session, which creates and opens one in a single call.
project_describe
ChatGPTSet what a project is for: description (one line, '' clears) and tags (a few lowercase words; replaces the list, [] clears). Requires ifRevision.
project_docs
ChatGPTThe Forge Projects reference, returned as markdown: using a project in a codebase (project_find, the skills to read, project_pull, adapting), remixing projects and collections, skills, the Vite-shaped tree (components, shared files, assets, package.json, index.css), code sessions (download, edit, upload, submit), limits and path rules, the story body, revisions and restore, publishing, error codes. Older tools in this server say fragment for a project and variant for a component; it is the same data.
project_feedback_list
ChatGPTThe feedback people left on the project's preview: comments pinned to a spot, proposed style edits, screenshots with a note — each with a reply thread and open/resolved status, and the componentId it was left on. Check it when you start and after you submit; answer with project_feedback_reply and close what you handled with project_feedback_resolve.
project_feedback_reply
ChatGPTAnswer a feedback item in its thread (what you changed, or a question back).
project_feedback_resolve
ChatGPTMark a feedback item resolved once handled, or reopen it (reopen: true).
project_find
ChatGPTSay what you want to build (intent) and, optionally, describe the codebase you are in (codebase: framework, styling, dependencies, notes). Ranks the projects across the user's workspaces and names the skills to read first. If gate.fits is false nothing fits well enough: build it yourself, still following the relevant skills. Then skill_get each skill and project_pull the one you chose.
project_get
ChatGPTEverything about one project except file contents: summary and revision, coverComponentId (the cover: shown first, the preview image, what share and publish links open), components (meta, story, entryPath), the projected tree, package.json dependencies, public state. Read before editing; pass its revision as ifRevision on your next write.
project_get_file
ChatGPTOne file of the project's tree. Text files come back inline as content; binary assets answer downloadUrl and deliveryPath instead. Read a file before you rewrite it.
project_list
ChatGPTThe projects in a workspace (name, description, tags, ids, revision, public state, sizes, collectionId) and its collections (id, name, projectCount). Scan description and tags to find the one the user means. orgId is only needed when you belong to several.
project_list_revisions
ChatGPTHistory by project revision, newest first: which tree paths each revision changed. With path, only that file.
project_move
ChatGPTFile a project under a collection of its workspace, or at the workspace root (collectionId: null). Requires ifRevision.
project_preview
ChatGPTWhen the user wants to see the project: answers shellUrl — the workbench page, the URL to hand them; component (id or name) opens that component. Also answers url, the raw sandboxed render for your own checks (never give the user that one), plus the project's name and components.
project_publish
ChatGPTPublish the project's current revision to its public URL (slug, claimed on the first publish), optionally password-protected. Tell the user what goes public, then call again with confirm: true. Requires ifRevision.
project_pull
ChatGPTOne read with everything needed to use a project in the codebase you are in: its files (inline, or a tar URL for big trees), signed asset URLs, dependencies, the component's story (props and defaults; component picks one by id or name, the cover by default), the governing skills in full, the file trees of its collection and workspace, and provenance (put provenance.header atop the adapted entry file). Changes nothing in the cloud.
project_remix
ChatGPTCopy a project into a workspace (orgId; the source's own by default), optionally renamed (name), filed under a collection (collectionId, null for the root), or just one of its components (componentId). Works for a project shared with a remix link.
project_rename
ChatGPTRename a project. Requires ifRevision.
project_restore_revision
ChatGPTRestore every component and file (or one path) to its state at a revision, written as a NEW revision. Requires ifRevision (the project's current revision).
project_set_cover
ChatGPTMake a component the project's cover: the one shown first, used for its preview image, and what its share and publish links open. Requires ifRevision.
project_set_password
ChatGPTPassword-protect the public page (one shared password visitors type once), or pass null to open it again. Takes effect at once. Requires ifRevision.
project_start_session
ChatGPTCall this FIRST for any confirmed create or edit, before research or drafting. Without projectId it creates the project (name required; pass a one-line description and a few tags too) and its first component. Answers shellUrl (the project's workbench page: live preview, controls and the agent-activity indicator — hand it to the user right away), workdir, the file list, download.command (a gzipped tar of the Vite-shaped project), upload.command, the session deadline and the project's revision. Run download.command, edit under workdir, run upload.command exactly as returned, then project_submit with projectId, sessionId and ifRevision set to the revision this answered. Components are src/variants/<slug>.tsx (the path keeps its old name). Never start a session for questions or read-only work.
project_submit
ChatGPTArgs: projectId, sessionId, ifRevision (the revision project_start_session answered), optional deleted (tree paths to remove) and component (id or name the URLs open). Applies the tree you uploaded, writing only what changed, in one revision; answers the new revision, shellUrl (the ONLY URL to give the user), previewUrl (for your own checks), what was written, and ignoredPaths. A rejected submit keeps the session: fix, re-upload, submit again.
project_unarchive
ChatGPTBring an archived project back, with its public page and share link. Requires ifRevision.
project_unpublish
ChatGPTTake the project's public page offline (the slug is kept for a later publish). Tell the user, then call again with confirm: true. Requires ifRevision.
skill_delete
ChatGPTDelete a skill and its history. Asks the user first: without confirm: true it answers confirmation_required. A workspace skill (no collectionId) needs an admin; a collection skill needs a member. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
skill_get
ChatGPTRead a skill: its fields, markdown (the rendered SKILL.md) and path (where it goes in a project: .agents/skills/<name>/SKILL.md).
skill_list
ChatGPTList a workspace's skills without bodies: every skill, one collection's (collectionId), or the workspace's own (collectionId: null). Read one with skill_get.
skill_list_revisions
ChatGPTA skill's history, newest first: revision, name, description, byteSize and author. Restore one with skill_restore.
skill_restore
ChatGPTRestore a skill to a past revision, as a new revision (the counter never rewinds). Answers the skill without its body. A workspace skill (no collectionId) needs an admin; a collection skill needs a member. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
skill_save
ChatGPTCreate or update a skill. Without skillId: creates — pass name, description and body, or markdown (a whole SKILL.md whose frontmatter supplies name and description); collectionId files it under a collection, omitted files it at the workspace root. With skillId: updates the given fields. Explicit fields win over markdown. Answers the skill without its body, plus the new revision. A workspace skill (no collectionId) needs an admin; a collection skill needs a member. Optional ifRevision refuses to overwrite a concurrent edit (revision_conflict carries the current revision).
forge_create
Claudeforge_delete
Claudeforge_read
Claudeforge_update
Claude