Quickstart

1. Sign in and get a key

Sign in to the Hunta console. Every account gets its own isolated memory store (the Solo tier is free for one seat, no card).

Mint a key from the Keys page. Two kinds matter:

• An agent key (memory:capture + memory:read) is scoped to propose and recall. An agent holding it can suggest memories and read sealed facts, but can never write canon directly. This is the key you give to harnesses. It writes through POST /v1/memories/capture (or the capture MCP tool); POST /v1/memories refuses it with 403 insufficient_scope.
• An owner key (memory:write) writes canon directly through POST /v1/memories (or the gather MCP tool). Keep it for yourself and for trusted tooling.

2. Write and recall

export GATHER_URL=https://mcp.hunta.ai
export GATHER_TOKEN=<your-key>

# owner key: write a memory straight to canon
curl -sS -X POST "$GATHER_URL/v1/memories" \
  -H "Authorization: Bearer $GATHER_TOKEN" -H "Content-Type: application/json" \
  -d '{"text":"Our design partner is Acme Corp."}'

# agent key: propose a memory (lands in the review queue, returns a candidate_id)
curl -sS -X POST "$GATHER_URL/v1/memories/capture" \
  -H "Authorization: Bearer $GATHER_TOKEN" -H "Content-Type: application/json" \
  -d '{"text":"Our design partner is Acme Corp."}'

# recall it (sealed facts only, with provenance)
curl -sS -X POST "$GATHER_URL/v1/memories/search" \
  -H "Authorization: Bearer $GATHER_TOKEN" -H "Content-Type: application/json" \
  -d '{"query":"design partner","limit":5}'

A capture lands as a candidate in your staging queue instead of going straight to canon. You review and promote candidates in the console (Queue), or with POST /v1/candidates/{id}/promote using an owner or curator key. Recall never returns unreviewed candidates, so the capture above will not show up in the search until it is promoted.

3. Prove the isolation

curl -sS -X POST "$GATHER_URL/v1/admin/verify-isolation?nonce=my-first-proof" \
  -H "Authorization: Bearer $GATHER_TOKEN" > receipt.json

npx -y @hunta/verify receipt.json

This runs live adversarial probes against your tenant boundary and returns an Ed25519-signed receipt, then verifies it against Hunta's published key. See Verify the attestation for what the probes mean and how the trust model works.

Next

• Connect over MCP so your agents use memory as a native tool.
• Install the Claude Code plugin for auto-recall on every prompt.
• Create a Tribe to share one attested store with your team.