Agents & MCP

An agent adds one URL and starts: the MCP server https://mithril.fund/mcp (streamable HTTP). It reads /mcp.json (a machine-readable map of the service), /llms.txt and /llms-full.txt (the docs, summarized and in full), /agent-quickstart.md (the steps). All public, no auth.

Start in one command

# Claude Code (public tools work without the token; inference needs it)
claude mcp add --transport http kotoba-cloud https://mithril.fund/mcp --header "Authorization: Bearer $KOTOBA_CLOUD_API_KEY"
# Codex CLI
codex mcp add kotoba-cloud --url https://mithril.fund/mcp --bearer-token-env-var KOTOBA_CLOUD_API_KEY
# Cursor / VS Code (.cursor/mcp.json, .vscode/mcp.json)
{"mcpServers":{"kotoba-cloud":{"type":"http","url":"https://mithril.fund/mcp","headers":{"Authorization":"Bearer ${KOTOBA_CLOUD_API_KEY}"}}}}
# the graph database is its own server: https://graph.kotoba.cloud/mcp (docs: /graph/)

Shell

curl -s https://mithril.fund/.well-known/mcp.json | jq '.tools, .clients'
curl -s https://mithril.fund/mcp.json | jq '.servers, .tools[].name'
curl -s https://mithril.fund/llms.txt
curl -s https://mithril.fund/agent-quickstart.md

/mcp and /mcp.json

/mcp is the MCP protocol endpoint (JSON-RPC initialize / tools/list / tools/call over streamable HTTP, stateless). Tools: kotoba_knowledge_search / kotoba_knowledge_context / kotoba_models (no credential) and kotoba_research_status / kotoba_chat_completion (the connection's Authorization: Bearer kc_pat_…). A tool result is the corresponding HTTP route's own answer — a 401 stays a 401.
/mcp.json is a configuration file that says how to call plain HTTP — each tool's inputSchema and URL — for an agent that does not speak MCP to wrap as its own tool.
public: /v1/knowledge/search?q= and /v1/knowledge/context?id= — read-only public security evidence (ATT&CK and others). CORS-open, no auth.
research: the research API's endpoints and sessionSetup (issue a PAT, start verification, a human completes it).

The agent's steps

The account holder issues a PAT at console.mithril.fund/account and puts it in the agent's environment (an agent cannot sign in).
Read GET /v1/research/status: eligible, or which reason. On verification-required, ask the person to verify (POST /v1/research/ekyc/start only returns a URL; a person completes it).
POST /v1/chat/completions (OpenAI shape, tools allowed); poll the job API for long work.
Run tools on the agent's machine; the server executes nothing (the AUP's local-execution boundary).

Claude Code / hermes and friends

# per-tool copy-paste: https://docs.mithril.fund/integrations/
# Anthropic-protocol clients need a translating proxy (docs: Anthropic SDK)

/mcp.json's tools: kotoba_knowledge_search / kotoba_knowledge_context (public), kotoba_research_run (the job API), kotoba_chat_completion (the OpenAI shape). The last two take a PAT bearer or a browser session and need a verified principal.