A production MCP server, already built
You could spend three weeks writing your own content MCP — handling content addressing, pre-signed URLs, approval state, multipart upload, and auth. Or you could plug into clipsilo and ship tomorrow. Built by engineers who got tired of solving the same plumbing problems.
4 tools out of the box · API-key auth · JSON-RPC 2.0 · OpenAPI 3.1 parity
4
Tools included
JSON-RPC
2.0 transport
60s
To first call
0
Servers to run
Everything your agent needs, day one
No scaffolding. No "this feature coming soon." A real, running MCP server your agent can talk to in the next ten minutes.
Available tools
search_assets
Full-text + faceted search across approved assets. Tags, content type, date range, title.
get_asset
Fetch full asset detail by ID, including pre-signed S3 URL and all metadata.
list_batches
Enumerate batch uploads with status filters (pending, approved, distributed).
get_batch
Retrieve a batch manifest with every asset, metadata, and approval state.
Transport
JSON-RPC 2.0 over HTTPS
Auth model
API key header (read or write)
Schema
JSON Schema on every tool
How the MCP endpoint works
Transparent, predictable, debuggable. Built the way you\'d build it if you had the time.
Per-workspace MCP URL
Each clipsilo workspace gets its own endpoint: https://mcp.clipsilo.com/v1/<workspace-id>. Request tools/list and you get back a JSON-Schema description of every available tool. Your agent runtime binds them automatically.
API key auth
Every tool call carries an Authorization: Bearer <key> header. Keys are scoped read-only or read-write. Rotation is instant — revoke in the admin panel, calls fail immediately. No OAuth dance, no token refresh loops, no stale credentials.
Response format
Every response is a JSON object with consistent envelope fields: results, cursor, total_count. Asset objects include content hash, pre-signed URL (default 1-hour TTL), approval state, and metadata. Deterministic, versioned, LLM-friendly.
Approval-gated responses
By default, the MCP server only surfaces assets in the approved state. A read key cannot retrieve drafts or rejected content. This is the safety primitive that makes agentic distribution workflows possible — autonomous agents physically cannot access un-approved bytes.
A real MCP tool call
Request in, response out. No magic, no hidden state.
Request — JSON-RPC 2.0
POST /v1/ws_7f2b9c/mcp HTTP/1.1
Host: mcp.clipsilo.com
Authorization: Bearer csk_live_...
Content-Type: application/json
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "get_asset",
"arguments": {
"id": "ast_7f2b9c"
}
}
}Response — approved asset
HTTP/1.1 200 OK
Content-Type: application/json
{
"jsonrpc": "2.0",
"id": 1,
"result": {
"id": "ast_7f2b9c",
"title": "Golden hour aerial",
"content_type": "video",
"mime_type": "video/mp4",
"bytes": 84328923,
"sha256": "a7f9c3...c1d4",
"approval_state": "approved",
"approved_at": "2026-03-08T14:22Z",
"tags": ["aerial", "sunset"],
"signed_url": "https://s3.amazonaws.com/...",
"expires_at": "2026-04-16T15:22Z"
}
}What developers build with it
Real agent workflows that ship because the content layer isn't a side project.
Agent-assisted social posting
Claude drafts the caption, pulls the matching clip, hands off to Buffer. No human in the distribution loop — only in the approval loop.
Automated content retrieval
RAG pipelines that surface relevant footage for a script, deck, or newsletter — with content hashes that make caching trivial.
Creative brief lookup
Agent reads a brief, queries the warehouse for on-brand visuals, returns candidate assets with approval state intact.
Approval-gated distribution
Autonomous workflows where a human sign-off is a hard prerequisite before any bytes leave the warehouse. Safety by default.
Why not build your own?
Nothing here is unique magic. But each piece takes real engineering time. Here\'s the honest breakdown.
Content addressing is hard
Streaming SHA-256 during upload without spiking memory. Handling multipart hash stitching. Writing a content-addressable store that\'s both fast to search and consistent across retries. Every time we\'ve seen a team try this, they underestimate it by 3x.
Approval workflows are hard
Single-use preview tokens. Expiration rules. Comment threads tied to asset versions. Audit logs that hold up in a compliance review. Notifications when a client actually opens the link. The "happy path" takes a weekend; the full feature takes a month.
S3 plumbing is boring
Pre-signed URLs with correct expiration. Multipart initiation, chunk signing, completion, abort. CORS configs that don\'t break on Safari. Lifecycle policies. The stuff you\'d rather not be writing when you could be shipping agent features.
If your product is the agent, make clipsilo the warehouse. If your product is the warehouse, we\'d love to hear what we\'re missing.
Questions MCP developers ask
The honest answers to what AI tooling engineers want to know.
Do you support the full MCP spec (resources, prompts, tools)?
Tools: yes, fully. Resources: partial — asset listings are exposed as resources for agents that prefer that discovery model. Prompts: not yet — we don't think the warehouse is the right place to store agent prompts. We track the spec closely and will ship new capabilities as they stabilize.
Can I extend the MCP server with custom tools?
Not directly — the MCP server is managed. But you can build agent-side tools that compose our four primitives. For example, a custom `get_approved_hero_shot(brand, aspect)` tool in your agent runtime can translate to a `search_assets` call under the hood. Keeps our server simple; keeps your agent expressive.
What's the latency profile?
P50 for search and get: under 80ms from our US region. Pre-signed URL generation is a few milliseconds. Large list calls paginate at 100 items per page. If your agent is hitting us in a hot loop, enable response caching on your side — hashes make cache invalidation trivial.
How do I test without a real workspace?
Sign up for the free trial — seven days, no credit card. The trial workspace has all MCP capabilities, including sample assets we pre-populate so you can immediately run `search_assets` and get real results back. Prefer a local mock? We publish an OpenAPI spec you can stub against.
What about rate limits and quotas?
1,000 requests per minute per workspace by default, bursting to 3,000. MCP calls share the same pool as REST. Response includes standard `X-RateLimit-*` headers. Above that, talk to us — we have a higher-volume tier for production agent pipelines.
Skip the plumbing, ship the agent
Start your 7-day free trial. Your MCP URL is provisioned the moment you sign up.
7-day free trial · No credit card · Cancel anytime