mcp-film
The mcp.film directory as an MCP server. Give your agent a
map of every MCP server worth knowing in AI filmmaking β searchable by
capability, with copy-paste install configs and a full-pipeline planner.
claude mcp add mcp-film -- npx -y mcp-film
Or connect the hosted endpoint β same tools, no local install:
claude mcp add --transport http mcp-film https://mcp.film/mcp
Speaks MCP 2025-11-25, 2025-06-18 and 2025-03-26 (an unknown offered version
gets the latest). initialize returns short instructions: start with
recommend_film_mcps(brief).
| Tool | What it does |
|---|
recommend_film_mcps | Match a free-text brief to the best routes and first servers to connect. Hard constraints first (constraints object or phrases in the brief), then relevance; confidence says how well the brief matched |
search_film_mcps | Tokenised, alias-aware search (lipsync = lip-sync = lip sync; tts = text to speech; 4k = upscale; creative-app names and aliases such as c4d, resolve or fcp find the servers that drive that app); filter by category, capability, official, remote, answering_only (hosted endpoint answered the daily check); paged |
get_film_mcp | Full entry for one server: install, auth, tools, caveats, status, evidence level and last endpoint/package check |
list_film_categories | All categories with pipeline stages and agent hints |
list_film_capabilities | Capability tags (aliases folded in) ranked by server count; min_count 2 by default; paged |
get_film_capability | Ranked server cluster for one tag; aliases resolve (lipsync β lip-sync, says resolved_from); paged |
get_install_config | Copy-paste config or hosted remote URL for a client from list_client_profiles; a bad client returns valid_clients |
list_client_profiles | Supported MCP client setup profiles and install-shape hints |
get_martini_setup | Executable runbook for connecting a user to Martini (the sponsor): resumable state model, human steps, agent steps per client / shell / browser, quote-first generation contract, structured cost, when not to |
get_catalog_pulse | Operating pulse: freshness, verification queue, curator agenda, machine surfaces. demand_signals and martini_growth_checks are curator/ops sections served only by name |
list_film_playbooks | Curated production stack recipes for common AI filmmaking jobs; paged |
get_film_playbook | Full playbook with primary servers, setup order, auth requirements, steps, constraints, fallbacks. Pass client (claude_code, claude_desktop, cursor) for only a copy-paste install of the whole stack |
list_film_recommendations | Intent-routed MCP shortlists for common filmmaking jobs; paged |
get_film_recommendation | Full recommendation with roles, reasons, fallbacks, and Martini handoff guidance |
plan_film_stack | Recommended servers for every pipeline stage, biased by your brief (capabilities, taglines, descriptions) |
submit_listing | Propose a new server: validated, deduped against the live registry, returned as a ready-to-file GitHub issue payload (a mcp.film maintainer verifies against primary sources before listing) |
Response shapes
Every result is compact JSON in content[0].text and the same object in
structuredContent.
Lists (search_film_mcps, get_film_capability,
list_film_capabilities, list_film_playbooks, list_film_recommendations)
carry count, total, limit, offset, next_offset and truncated.
Nothing is cut silently: pass offset: next_offset for the next page.
Default limit is 20 (50 for capabilities), maximum 100.
Compact server (search results, capability clusters, stack picks):
slug, url, verified, caveat, name, vendor, official, category, tagline, pricing, remote, evidence, capabilities plus status when the listing carries one.
caveat is the entry's caveats[0], else the first sentence of its notes
that is not a verification log line. get_film_mcp is always the full entry.
Recommendation pick (recommend_film_mcps): primary_servers[] and the
alias primary[] hold {slug, name, role, why, remote, official, verified, status?, caveats?}; fallback_servers[] hold {slug, name, remote, official, verified, status?}. Pass verbosity: "full" for the complete server objects.
Each route has match: {score, strength}; the reply has confidence
(strong | fair | weak). When Martini is a primary pick or the brief asks
for production state, the route carries martini_handoff (prose) and
martini_handoff_structured (when, not_when, first_calls, install {mcp, cli}, signup_url, setup_tool, tracked_url, disclosure).
Constraints (recommend_film_mcps): constraints: {local_only, no_upload, hosted_only, has_shell, free_only, os}, also inferred from phrases such as
"locally", "without uploading", "confidential", "offline", "no budget", "free
tools", "no local install". Violating picks are removed before scoring and
named in excluded_by_constraints; if nothing survives, the reply is
no_match with unsatisfied_constraints and relevant_routes.
Whole-stack install (get_film_playbook with client): {id, title, url, install_script, note}. install_script is {client, servers, script} for
claude_code (one commented claude mcp add line per server) or {client, servers, file, config, note} for claude_desktop / cursor (one merged
mcpServers object), plus checklist[] (slug, name, auth, env_var?, where?:
the keys, OAuth sign-ins and local apps the stack needs) and skipped[] (slug, name, reason, listing). Primary and workflow servers only, never fallbacks.
Retired listings, listings with install.generate: false, and listings with no
mechanical install for that client are skipped with the reason, never guessed
at. It is the same block shown on mcp.film/playbooks/{id}/ and in its .md.
Evidence. Every server carries evidence, derived at build time from a
daily anonymous check (no LLM, no credentials, tools are never called):
endpoint_answered (the hosted endpoint completed an MCP initialize, or
answered with a 401/403 sign-in challenge, within the last 7 days) >
package_resolves (its npm/PyPI package exists, checked within 7 days) >
docs_checked (only the curator's verified date). get_film_mcp adds
last_probe: {date, result, protocol_version, tool_count, anonymous_tools_list, auth_required, oauth_metadata, consecutive_failures, last_ok} (result: ok | auth_required | handshake_only | error |
unreachable) and package_check: {date, registry, name, found, latest_version, last_ok}, each null when not applicable. A failing check
never changes status; a curator does that after re-verifying.
Status. Listings with status dead, archived or deprecated are hidden
from search, recommendations and stack plans unless include_inactive: true.
Zero results. search_film_mcps returns did_you_mean (nearest
capabilities, categories, servers) and a hint; an unknown category or
capability is an error carrying valid_categories / did_you_mean.
How it works
Zero dependencies. Fetches the live registry from
https://mcp.film/api/registry.min.json and production playbooks from
https://mcp.film/api/playbooks.json, plus recommendation routes from
https://mcp.film/api/recommendations.json and the operating pulse from
https://mcp.film/api/pulse.json (5s timeout each), then falls back to bundled
snapshots offline. Live data is cached for an hour and refreshed in the
background (MCP_FILM_CACHE_TTL_MS); MCP_FILM_OFFLINE=1 serves the snapshots
only and MCP_FILM_API_BASE points the fetch at another host.
get_catalog_pulse reports data_source and observed_at. Data is curated and
re-verified continuously at
github.com/c47-inc/mcp-film.
MIT.