ArkHive (memory) and Algernon (parallel agents) are free and work on their own, but they work best under Sentarion, which checks every task before it runs and records it. Sentarion 2 keeps a fleet from overspending: budget and token caps, cached results so you never pay twice, and retries for failed tasks. The Claude Code Seatbelt Kit is $27 once and includes a 14-day Sentarion 2 trial.
Claude Code writes fast. It also forgets the plan between sessions, rm -rfs the wrong folder, pastes a key into a commit, restarts a live service and says "done" without running the code once. The Seatbelt is a set of deterministic hooks that sit underneath the agent, so it does not matter whether the model remembers to be careful.
Five policies: never delete production data · never expose credentials · never change deployment without approval · verify frontend-to-backend wiring · test before declaring completion. Refused or sent to you for a yes, with a plain-English reason.
Every session opens with a brief: what was edited, run, failed and decided. Tamper-evident local chain; nothing leaves your machine.
The agent cannot end a session with untested edits, or with a screen calling routes no backend defines. Plus /fix-my-app, /finish-end-to-end, /audit-real-vs-stubbed.
pip install sentarion-mcp then sentarion seatbelt install).
No card. We email nothing automatically; the key appears right here. We do keep your address so a human can follow up.
ArkHive and Algernon run on their own, but they are built to work together under Sentarion. It checks every task before it runs and logs it to both memory chains. When you want spend caps, cached results and signed records, that is Sentarion 2.
Tamper-evident local memory. Your AI writes what it did and reads it back next session; every record is hash-chained, so a silent rewrite fails verification.
pip install arkhive-mcp
Tools: birth · remember · recall · verify · govern
Fan a goal out to cheap parallel workers on your own key, or on a free local Ollama it detects automatically. The expensive mind stops doing the grunt work.
pip install algernon-mcp
Tools: algernon_doctor · plan · dispatch · orchestrate
{ "mcpServers": {
"sentarion": { "command": "sentarion", "env": { "ANTHROPIC_API_KEY": "sk-ant-..." } },
"arkhive": { "command": "arkhive-mcp" },
"algernon": { "command": "algernon-mcp" }
} }
No key? Leave it out. Algernon and Sentarion use your local Ollama when one is running. Hosted, no install: https://arkhive.dondatabrain.com/mcp and https://arkhive.dondatabrain.com/sentarion/mcp.
Fan work out to a fleet and the bill adds up. With Sentarion the expensive model only plans and you talk to it, while the grunt work runs on cheap workers or a free local model. Sentarion 2 sets a budget and token caps, caches results so you never pay twice for the same answer, and retries failed tasks. Everything in the free line stays free.
| You need to… | Free 0.x | Sentarion 2 |
|---|---|---|
| Start a session already knowing what happened | page through recall | context_pack: one call, fits your token budget |
| Find the record about the nginx deploy from last month | last N rows only | full-text search, actor/kind/tag/time filters, paging |
| Keep client A's memory away from client B | one global chain | spaces per project/tenant; hosted keys pinned to a space |
| Prove nothing was altered, fast | re-hash everything | incremental + Ed25519-signed verify that names the bad block |
| Stop the agent before it emails a customer list | only flags you remembered to declare | inferred risk flags (PII, credentials, money, irreversible, outbound, bulk) + stored policies + REVIEW verdict |
| Run 40 sub-tasks without the tool call dying | dies at ~120 s | progress notifications + background jobs |
| Cap spend and never pay twice | — | budget_usd, token caps, result cache, retries |
| Feed task B the output of task A | ordering only | {{id}} data flow between tasks |
| Prove what the fleet actually did | goal string logged | signed manifest: prompt/result hashes, usage, cost |
| Review a diff adversarially before merging | — | multi-agent review: finders per dimension, skeptics refute |
| Let the fleet edit code safely | worktree create/remove | sandbox: create · diff · apply_patch · commit |
The full v2 control plane for you alone: task contracts, phase gates, an agent that cannot approve its own work, SHA-bound evidence, signed receipts, budgets. One licence seat, one hosted tenant. Try it free for 14 days with the trial key above, then cancel anytime.
Subscribe — $29/moThe full InboxAxe platform: replies and posts drafted for you, and nothing sends without your yes. The same plan as on inboxaxe.com.
Get started risk free: $27 for your whole first month, then only $77/mo. Cancel any time — email us and we confirm by reply.
Get started risk free — $27More than one developer or client? Tell us what you need and we will send you a price.
Ask usHand-delivered install and pilot on your own server, unlimited v2, your own hosted endpoint.
Contact usAlready a Pro customer? Sign in → Settings → MCP to mint your v2 key.