| What it does | High-assurance infrastructure for AI agents
| An engine for long-running agents that rebuilds its prompt under a cap and runs on your budget
|
|---|
| Stage | Not given
| MVP
|
|---|
| Founded | Not given
| Not given
|
|---|
| Based in | US
| Not given
|
|---|
| Pricing model | Open source
| Freemium
|
|---|
| Pricing | Not given
| Prototype: free during beta, then from December 16, 2026 it costs $1 per 1,000 agent calls after 1,000 monthly included calls. Production: $20 a month with 20,000 agent calls included, then $1 per 1,000 calls after that. Top-ups: $1 per 1,000 calls, from $10 to $500 at a time, never expiring.
|
|---|
| Key features | - Harpe: chooses what an AI agent reads under a token budget with consistent results
- Kibisis: keeps agent memory on your hardware with AES-256-GCM encryption
- Argus: append-only record that detects edits with cryptographic verification
- No network code by default for local and disconnected use
- MIT License with published code, benchmarks and raw results
| - Prompt rebuilt from journal under token cap at each step to keep it bounded as history grows
- Spending limits checked before each model call at token prices you declare
- Resume after crash without repeating finished actions
- Tool calls returned as actions for your code to carry out
- Proof of what ran: check prompts, re-run engine, download hash-chained journal
- Integrations with Vercel AI SDK, OpenAI Agents SDK, LangGraph, Mastra, Pydantic AI
- MCP server for Claude Code, Claude Projects and Codex
- Works with any model via bring-your-own-key
- TypeScript and Python clients for HTTP API access
|
|---|
| What makes it different | - Measured empirical validation across 6,111 questions outperforming full-context models
- Zero plaintext on disk with encrypted vaults matching plaintext accuracy
- 78.3% fewer tokens sent to the model with a 1,200-token budget per turn
- Preserves conversational chronology and relational qualifiers under token budgets
| - Bounded prompt under a fixed cap while history grows, so costs per call stop climbing on long sessions
- Every run recorded with cryptographic proof: rebuilable prompts, hash-chained journal, re-runnable engine
- Resumes from crash without repeating finished actions using stable idempotency keys
- Budget enforcement before each model call, not reactive overage billing
- No model key storage or logging; your provider bills you directly
|
|---|
| Free plan | Not given
| Not given
|
|---|
| Open source | Not given
| Not given
|
|---|
| Platforms | Not given
| Not given
|
|---|
| Public API | Not given
| Not given
|
|---|
| Community votes | 0 | 0 |
|---|
| DR (Domain Rating by Ahrefs) | 42 ▲ +1 (up) since last check | Not checked yet ● no change since last check |
|---|
| Trust Flow (Majestic) | 3 | 0 |
|---|
| Citation Flow (Majestic) | 24 | 9 |
|---|
| Referring domains (Majestic) | 62 | 4 |
|---|
Perseus Computing
Argentic