Prism

Prism

Stateless MCP gateway. Sits between an AI client and an MCP server, relays JSON-RPC verbatim, and turns every call into an OpenTelemetry span — shipped over OTLP to any endpoint you choose. No database, no accounts, no provisioning. See docs/PRISM.md for the full spec.

Fastest path to a live dashboard: get a free key at telinea.terradev.cloud, set one destination, and every span appears in Telinea within seconds.

Run

pip install -r requirements.txt

Zero-config dev: spans print to stdout.

PRISM_UPSTREAM_ALLOW='*' \ PRISM_ALLOW_INSECURE_HOSTS='localhost,127.0.0.1' \ uvicorn prism.main:app --port 8080

Where spans go

One pipeline, one egress primitive: PrismSpan → Serializer → Dispatcher. Prism knows four destination schemes and nothing about backends:

• `https://` / `http://` — POST serialized spans to any OTLP endpoint. Grafana, Honeycomb, a collector, Telinea — read the backend's own docs for its endpoint and auth header, fill in two fields, done.
• `file://` — append JSONL to a local path. No account, no network.
• `stdout://` — human-readable lines for development.
• `multi://a,b` — fan-out to named destinations; failures are isolated.

Configure in ~/.prism/config.toml (or ./prism.toml, or PRISM_CONFIG):

[prism]

destination = "telinea"

[destinations.telinea] endpoint = "https://api.terradev.cloud/telinea/v1/ingest" format = "telinea" headers = { "Authorization" = "Bearer tl_your_key" }

Formats: otlp_json (default — the industry wire standard), otlp_proto (needs pip install opentelemetry-proto), jsonl, telinea (native envelope for Telinea ingest).

Every field has an env override — secrets never need to touch disk:

PRISM_DESTINATION=telinea

PRISM_DESTINATION_TELINEA_HEADERS_AUTHORIZATION="Bearer tl_..."

Hosted mode: "Authorization" = "Bearer {token}" resolves {token} to the caller's bearer per span — each user's spans land in their own account.

Delivery: buffered, flushed every 5 s or 512 spans; Retry-After honored on 429 for every backend; failed batches go to a dead-letter ring buffer (10k spans) retried on backoff; on shutdown, in-flight spans drain and the dead-letter buffer is written to ~/.prism/deadletter.jsonl and replayed on next start. Fire-and-forget — Prism never blocks a chat session on a flush.

Client config

The entire integration is a URL swap plus headers:

{

"mcpServers": { "acme-prod": { "url": "https://prism-mcp.terradev.cloud/mcp", "headers": { "Authorization": "Bearer tln_live_…", "X-Prism-Upstream": "http://localhost:9000/mcp", "X-Prism-Project": "acme-prod", "X-Prism-Env": "development" } } } }

• X-Prism-Upstream — the real MCP server. Must match PRISM_UPSTREAM_ALLOW.
• X-Prism-Upstream-Authorization — forwarded upstream as Authorization.
• X-Prism-Capture — none (default) | args | args,results. Redacted.
• X-Prism-Sample — 0.0–1.0, sampled once per session.

Header-less clients can encode the upstream in the path: POST /mcp/u/{base64url(upstream)}.

Registries (Smithery et al.)

Registry installs can't set custom headers — config arrives as query params. Every X-Prism-* option has a query-param twin:

/mcp?upstream=https://mcp.acme.dev/mcp&api_key=tln_live_…&project=acme&env=prod

• upstream → X-Prism-Upstream (or set PRISM_DEFAULT_UPSTREAM for single-upstream deployments)
• api_key / token → Authorization / X-Api-Key
• upstream_authorization → X-Prism-Upstream-Authorization
• project, env, role, capture, sample → their X-Prism-* twins

smithery.yaml in this directory is the registry manifest for the self-hosted container path; the hosted listing just points at https://prism-mcp.terradev.cloud/mcp.

What you get

One MCP session = one trace. One JSON-RPC request = one span. Tool failures surface correctly (result.isError → ERROR, not a green 200). With a Telinea destination, every Prism-fronted server also appears on the Telinea Gateway board via the endpoint attribute — p50/p95/p99, req/min, sparkline — with no backend changes.

Not built (by design)

• No backend-specific serialization. If a backend needs a proprietary format, run an OpenTelemetry Collector and point Prism at it — that's what the Collector is for.
• No query layer, storage, or UI. Prism emits; Telinea stores and visualizes.
• Legacy HTTP+SSE transport (/sse + /messages, 2024-11-05 spec).
• DNS-rebinding-safe IP pinning at the transport layer (see guard.py).