Connect your AI to the fund.
New Era is a hosted MCP server. You bring your own AI, wire the councils you need, and get the fund's institutional method back through one secure connection. Your AI does the synthesis on your data — the fund runs no model and stores nothing.
What this is
New Era exposes a single externally reachable door: a JSON-RPC 2.0 MCP endpoint at POST /mcp. You connect your own AI client (BYOK — bring your own key), pick the councils that match your work, and your AI calls their tools. Each call returns the fund's institutional method — the framework, the stop-the-line gates, what "good" looks like, the exact output shape — applied as a lens to the task you pass in. Your AI then produces the finished work using your data, on your side. The fund runs no LLM of its own and never stores your inputs; a mandatory outbound filter strips any sensitive detail before anything crosses the connection.
Quick start
From zero to your first institutional answer in four steps.
-
Create an account & generate an API key
Enter your room, then click generate key under "Connect your AI". The key is shown once — copy it immediately. It looks like
neif_…and travels as a Bearer token. The gateway only ever stores a hash of it, so a lost key is rotated, never recovered. -
Add the MCP server to your AI client
In your MCP-capable client (Claude Desktop, or any other), open Settings → Developer → Edit Config and paste this block. Swap in your own host and key:
{ "mcpServers": { "newera": { "url": "https://YOUR-APP-HOST/mcp", "headers": { "Authorization": "Bearer YOUR_KEY" } } } }On neifcapital.com the URL ishttps://neifcapital.com/mcp. Self-hosted? Use your own deployment host, e.g.https://your-app.fly.dev/mcp. Restart your client after saving. -
Your AI discovers the tools
On connect, your client calls
initializeand thentools/list. The gateway returns the always-on public tools, plus the tools of every council your account owns, plus your private-KB tools if you're on the hosted-KB plan — and nothing else. Each tool's description includes its credit cost. -
Ask a question that matches a wired council
Describe your task in plain language; your AI picks the right tool and calls it via
tools/call. Start with something the public corpus answers (search_knowledge), or — once you've wired a council — paste a real brief and ask for the method (e.g.eval.fullfor a full company evaluation,quant.varfor portfolio risk).
The tools
A narrow, metered, audited surface. Read / analyse-only. No tool reads another tenant's private data.
Public tools — always on
These three are available to every account and read only the fund's published, public corpus.
| Tool | What it does | Input | Cost |
|---|---|---|---|
| plan.work | The best first call. Describe your task in plain language and it returns which councils and tools the job needs, in order, plus a forecast of how many tokens it will take to run on your AI — and if it's a heavy job, a cheaper, scoped alternative to start with. It plans only; your AI runs the tools. | task — string: what you want done. |
1 cr |
| search_knowledge | Searches the fund's published research corpus and returns matching titles and summaries (with the entry id and topic). | query — string: what to search for; optional limit (1–25). |
8 cr |
| get_knowledge | Fetches the full text of one published research entry by its id (capped to keep your context lean). | id — integer: an entry id from a search result. |
5 cr |
| list_topics | Lists the public topics the New Era model can answer about — a quick map of the corpus. | none | 2 cr |
Council tools — council.method
Every council you own exposes its own tools, named council.method. A few examples across the bench:
Each council tool takes a single input — free text: your task plus the data the council needs to work on. Be concrete: paste the figures, statements or brief, or describe the company, deal, portfolio or question. The tool returns the fund's institutional method for that task — your own AI then applies it to your data. The fund runs no LLM and stores nothing; you get the framework and structure back, your AI does the synthesis.
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "eval.full",
"arguments": {
"input": "Evaluate Acme Co. for acquisition. Revenue $42M, EBITDA $7M, ..."
}
}
}
The complete catalogue — every council, its tools and their per-call credit costs — is on the Pricing page and inside your room. Council tools you don't own still appear to you only if you own them; calling one outside your plan is not an error, it returns a short note telling you to add that council in your room.
Knowledge base
Public corpus for everyone · your private KB stays yours.
There are two knowledge layers, and they never mix:
The public corpus is read-only to everyone, reached through search_knowledge / get_knowledge / list_topics. It exposes only rows marked public, from a table with no client columns.
Your private KB is per-tenant and fully isolated — scoped to your account, never shared, never blended with the public corpus or another tenant. If you're on the hosted-KB plan, two extra tools appear:
| Tool | What it does | Input | Cost |
|---|---|---|---|
| kb.search | Searches your private hosted knowledge base (your tenant only). | query — string: what to find in your KB. |
3 cr |
| kb.add | Proposes an entry for your private hosted KB. It lands as a PENDING proposal you approve in your room's Inbox (human-in-the-loop) — not searchable until you approve it. | title, body (required); topic (optional). |
2 cr |
kb.search / kb.add are not exposed by the hosted gateway at all.Set your jurisdiction
The model is worldwide and assumes no country — until you tell it yours.
New Era is jurisdiction-neutral by design. To ground every council in your market, run the jurisdiction.setup tool once (just ask your AI: "set my jurisdiction"). It walks a short questionnaire — your regulator, currency, registries, accounting standard and tax — and your AI saves the resulting profile into your own knowledge base. After that, every council grounds its analysis in that profile.
Billing
Credits, per call. You only pay for what runs.
Billing is metered in credits, charged per call. Every tool shows its credit cost up front (in this page's tables and in each tool's description via tools/list). You hold a credit balance; each metered call runs inside a credit escrow — if your balance can't cover the cost, the call is refused and nothing is returned or charged. Every metered call (ok, denied or insufficient) is written to an auditable ledger, so you can reconcile exactly what ran. There's no subscription lock to use a single council — wire what you need and top up your balance when it runs low.
See the full price list and credit packs on the Pricing page.
Errors
Standard JSON-RPC 2.0 error objects: { "jsonrpc": "2.0", "id": …, "error": { "code": …, "message": … } }
| Code | Meaning | What to do |
|---|---|---|
| -32001 | Unauthorized — missing or invalid API key. | Check the Authorization: Bearer … header. Regenerate your key in your room if needed. (HTTP 401.) |
| -32002 | Insufficient credits for the requested tool. | Top up your credit balance, then retry. Nothing was charged. |
| -32004 | Diversity throttle — too many different tools called in a short window (the bulk-export / model-harvest signature). | Method tools are for analysing your deals, not bulk export. Space out distinct-tool calls; normal use (repeating a few tools) is unaffected. |
| -32029 | Rate limited — too many requests, too fast. | Slow down and retry. Honour the Retry-After header (≈60s). (HTTP 429.) |
| -32601 | Method / tool not found — unknown JSON-RPC method, or a tool not on the allow-list. | Call tools/list to see exactly which tools your account can use, and call one of those. |
| -32602 | Invalid params — the tool rejected your arguments (e.g. kb.add with an empty title or body). |
Read the message; it states what's wrong. Fix the arguments and retry. (No charge — refunded.) |
| -32603 | Internal error — an unexpected failure on our side. | The message includes a request id. Quote that id to support so we can trace it in the logs. |
FAQ
Does the fund see my data?
No. You bring your own AI (BYOK) and it does the synthesis on your side. Council tools return method, never store your inputs, and a mandatory outbound filter strips client/LP/project names, emails, account-like digits and deal figures before anything crosses the connection.
Which AI clients work?
Any MCP-capable client — Claude Desktop or any other tool that speaks the Model Context Protocol. The gateway is plain JSON-RPC 2.0 over HTTP with a Bearer header, so anything that can add an MCP server with custom headers can connect.
Can I build my own agents?
Yes — that's the Agent Factory council. Use build.agent to draft a complete institutional-grade agent spec (role, method, gates, tools, self-eval) from your intent, then agent.evaluate to grade it 0–10 against the same rubric the fund holds its own agents to, and loop build → score → refine until it's good.
Can I export results as HTML/PDF?
Yes. Use build.document — it returns the institutional document method, and your own AI writes the finished HTML/PDF on your machine. The fund stores nothing; the artefact never leaves your side.
What if a council tool isn't in my plan?
It's not an error. The call returns a short note naming the council and telling you to add it in your room. Your tools/list only ever shows the tools you actually own, so your AI generally won't reach for one you can't use.
What's the uptime / SLA?
Availability and support targets are documented separately — see your plan terms or contact support. Sources degrade gracefully: if a live-data feed is down, a call returns a "source unavailable" note rather than failing.