MCP Connector
Connect Claude — and any MCP-compatible AI assistant — directly to the OCS interactive science tools via the Model Context Protocol.
Endpoint
The server is a stateless Cloudflare Worker running on the Streamable HTTP MCP transport (2025-11-25). All tool logic runs client-side in your browser — the MCP server only returns deep-links and metadata; zero data leaves your machine.
Connecting your AI assistant
The OCS endpoint uses the Streamable HTTP MCP transport (spec 2025-11-25) and works with any compliant client. Select your client below.
-
1Open the Claude desktop app and go to Settings → Connectors.
-
2Click Add connector and select Custom MCP server.
-
3Paste the endpoint URL:
https://mcp.omegacentauri.me/mcp -
4Name it Omega Centauri Society and save. Tools appear under Read-only connectors.
{"status":"ok"}. Verify the server is up before connecting.
Add to ~/.cursor/mcp.json (create it if absent). Cursor picks up changes on the next project reload.
~/.cursor/mcp.json applies globally; a per-project override lives at .cursor/mcp.json in the workspace root.VS Code 1.99+ native MCP — add to .vscode/mcp.json in your workspace:
Cline extension — open Cline settings → MCP Servers → Edit config and add:
Any HTTP client can call the endpoint directly. The tools/list method confirms connectivity:
text/event-stream.W5 — Agentic literature loop
W5 combines the OCS MCP with NASA ADS and an arXiv MCP in a single agent session. The agent fetches recent mass-constraint papers, propagates any new measurements through the OCS constraint chain, and outputs an updated evidence ledger with deeplinks.
Required MCP servers:
- 1 · OCS
https://mcp.omegacentauri.me/mcp— no setup beyond connection - 2 · NASA ADS github.com/prtc/nasa-ads-mcp — requires a free NASA ADS API key
- 3 · arXiv Any arXiv MCP — search the MCP registry for
arxivand follow that project's README
Three-server config skeleton (Cursor / VS Code — fill in the ADS and arXiv entries from each project's README):
Agentic loop prompt: the copy-paste prompt now lives in the Example Prompts Library as agentic-literature-loop. Connect the three servers above, open the library card, and copy the prompt from there; the library version is the maintained one.
Example prompts
Copy-paste agent prompts for common OCS research tasks now live in the Example Prompts Library: 30 prompts rendered from a machine-readable SSOT, each ending at a surface you can verify without trusting us. The recipes that used to sit on this page are superseded by these library cards:
- E2 find-the-right-tool · survey the catalog, run one tool, mint a deeplink that reproduces the run
- S2 run-the-evidence-window · the machine chain that turns ten published measurements into the allowed IMBH mass window, then breaks it
- V1 stack-the-constraints · all ten constraint channels with floor/ceiling provenance and page-versus-worker parity
- G1 agentic-literature-loop · the OCS + NASA ADS + arXiv loop (successor to the W5 block above)
The library also carries everyday tasks, by-domain briefs, education tracks, and all 42 catalog chains as reading paths; it supersedes every copy-paste recipe this page used to carry.
Available tools
The worker exposes 11 callable tools: 4 meta tools (search/deep-link/verify/execute) and 7 server-side compute kernels. Live-verified against mcp.omegacentauri.me/mcp tools/list. It also serves the Example Prompts Library over the MCP prompts capability: prompts/list returns the 30 verification prompts and prompts/get returns one in full.
Meta tools
#in= fragment that opens the tool pre-filled with your values.- query string — free-text search against title and description
- category string — filter:
imbh-evidence·bh-physics·fermi-paradox·fermi-seti·mth·kardashev - register string — filter:
peer-reviewed·speculative - limit number — max results (default 20)
fields per step to receive pre-filled URLs via the #in= fragment convention. These are narrative reading paths for a human or agent to follow, not server-executed pipelines; use run_chain for a server-executed pipeline instead.- chain string, exclusive with steps — name of a pre-defined chain (42 available, see below)
- steps array, exclusive with chain — ad-hoc list of
{tool, fields?}objects
{policy_parameters, output_payload} and compares it to the claimed execution_hash. Works on artifacts from any ChainGraph-conformant source, not only OCS tools.- artifact object, or the three fields below: a full ChainGraph artifact
- policy_parameters / output_payload / claimed_hash: alternative to
artifact
constraint-stacker, bayes-factor-router, jwst-accretion-ledger, gwtc-remnant-classifier, roman-microlensing) in sequence, not deep-links. Supports §21.4 decision gates: an RFC 6901 pointer into a step's output_payload is matched against ordered rules (eq/neq/gt/gte/lt/lte/in/present/absent) to route to the next step.- chain string, exclusive with steps: one of 3 machine-executable chains:
imbh-evidence-window(default) ·evidence-threshold-routing·accretion-eligibility-fastfail - steps array, exclusive with chain: ad-hoc ordered list of
{tool_id, fields?, gate?}objects
Compute kernels (7)
Each returns a hash-anchored OCS ChainGraph artifact JSON; each mirrors a client-side browser tool but also runs server-side for chaining via run_chain.
supports_null · weak · substantial · strong · decisive. Used as a §21.4 decision gate for model-selection chains.Named chains (42)
Pass any of these as the chain parameter to build_ocs_workflow_links.
Prefill-enabled tools
22 of the 35 catalog tools accept #in=<base64url(JSON)> URL fragments (verified against OCS_APPLY_PREFILL usage in the tool source, matching tools-manifest.json's _meta field). Pass values as a fields object in build_ocs_workflow_links steps to receive a pre-filled URL.
radio-seti — all 6 params (eirp, dist, aeff, tsys, tau, bw) are log₁₀. qpo-mass-spin — lognu is log₁₀(Hz). bz-kardashev — hash keys are spin (not a) and power (not P). Do not pass linear values to log₁₀ parameters.
Technical details
Transport: Streamable HTTP, MCP spec 2025-11-25. POST /mcp, JSON-RPC 2.0, stateless (no session IDs).
Runtime: Cloudflare Workers (free plan), deployed globally. Data fixture (tools-manifest.json) served via ASSETS binding, cached per isolate.
CORS: omegacentauri.me, claude.ai, app.claude.ai, and localhost are in the allowed-origins set.
Privacy: The MCP server returns only URLs and metadata. All physics calculations run client-side in your browser — no inputs or results pass through the server.
Source: github.com/PostOakLabs/OCS