Claude and MCP
Jepela ships as an MCP server and a Claude Code plugin, so Claude can make typed decisions with memory as tool calls.
Claude Code
export JEPELA_API_KEY=jj_live_...
export JEPELA_MODEL=jepela-english # optional: the model decide uses when a call names none
/plugin marketplace add jepelasweden/jepela
/plugin install jepela@jepela
The plugin brings the tools and a skill on writing questions and memory lines that work.
Tools
| Tool | Does |
|---|---|
decide | answers typed questions about a state; model, subject, compare (with a subject), robust, quality, values, derive |
remember | stores facts about a subject, with write-time warnings |
memory | what Jepela remembers about a subject, word for word |
forget | removes a concept and checks recall, and lists what it removed where the case did not show the name and what it kept that may be about the name; preview says what would go and removes nothing |
delete | removes a subject's whole memory, and the subject's requests from your batches |
feedback | the right answer for an earlier decision |
quality | accuracy and calibration from that feedback |
Each successful result has a short text for the model and the full API response as structuredContent (read by clients on MCP 2025-06-18); an error result has the text only. Windows, the cache, top_k, min_share, placebo, rules, golden sets, batches and calibration are not tools: use the API or an SDK for those. decide does fire the rules a subject already has.
Other MCP clients
The server is one file of standard-library Python (plugins/jepela/server/jepela_mcp.py) speaking MCP over stdio, versions 2025-06-18, 2025-03-26 (including batches of messages) and 2024-11-05:
{"mcpServers": {"jepela": {"command": "python3", "args": ["/path/to/plugins/jepela/server/jepela_mcp.py"],
"env": {"JEPELA_API_KEY": "jj_live_...", "JEPELA_MODEL": "jepela-english"}}}}
Without a key it still starts, and every tool call says how to get one.