persistent PTY
Shell session per user. cwd, env, open REPLs survive across tool calls in the same session.
Core capabilities in one self-contained binary, before you add skills or plugins. Free to use. Closed source by design. Runs on your machine.
Claims below are either reproducible with a shell command or documented on this site.
Hierarchical memory with hybrid recall. Product landing:smartopol.com/memory.
Raw moments, topic summaries, higher patterns. One query can search every level.
· keyword + meaning, scored together
· multi-language retrieval
· local file default · optional Postgres
Bounded job consolidates the day, updates a small profile, writes a dated receipt you can read and revert.
Personal memory stays personal. Shared index holds runbooks and household facts every user agent can recall.
Writes pass threat patterns. Recalled memory is fenced so old injection text cannot rewrite the system prompt.
Deep dive: /how-it-works#memory
Persistent PTY with cwd and env. Not a one-shot sh -c that forgets state every call.
Shell session per user. cwd, env, open REPLs survive across tool calls in the same session.
Drive vim, less, fzf, tmux, Python/Node REPL, and coding agent CLIs where configured.
Start, poll, wait, kill long builds and tests without blocking the main conversation.
Orchestrate Claude Code, Grok Build, or similar agents as tools. Smartopol stays the long-lived brain. Grok is a model; Grok Build is a coding agent product.
read_file · patch_file · write_filelist_files · search_files · glob_filesweb_search · web_fetchSame person on Discord, Telegram, web, and terminal can share one identity via pairing codes.
One-shot, recurring, or cron. Optional shell payload. Optional agent turn.
Scoped tools and budgets. Shared Memory Pyramid. Idempotency guards for long jobs.
Skills live under ~/.smartopol/skills/. No rebuild. Evergreen skills cannot be uninstalled by memory injection.
Any language that can speak the host interface. Capability manifest at load time. OS isolation. Signed third-party packages.
Price and distribute as you like. We do not take a marketplace cut.
Email hello@smartopol.ai with what you are building and what the SDK lacks.
Pick a primary, a fallback, a local last-resort. Change them without rebuilding the product.
Stable system prompt and slow-changing profile can pin to provider prefix cache when the provider supports it. Long sessions re-pay less for the same static tokens.
When configured, Smartopol can use the local claude CLI you already authenticated. That respects how Anthropic expects the subscription CLI to be used, instead of scraping tokens into third-party apps. Tradeoff: slower than raw API, cleaner terms and billing.
Per-session override: /model [name]. Shared history across failover when the router switches providers.
Threat model includes poisoned memory and hostile plugins, not only “trusted laptop.”
Encrypted vault for API keys. Master password you choose. Config references vault:name instead of plaintext keys on disk.
Roles and tool grants. Memory isolation between users is architectural.
Per-user windows so one guest cannot starve the box.
Parent on Grok/Claude; kid on local Ollama. Same service.
One bad session should not take down others. Inspect with smartopol sessions.smartopol audit verify walks the chain.Context usage visible before you blow the window.
List, inspect, kill without restarting the service.
Latency and rate-limit signals for the router.
Off by default. Opt-in levels only.
Structured JSON on stdout for your stack.
We do not claim HIPAA/SOC2 badges we have not earned. We claim local-by-default data path, audit log, exportable memory, and NDA source review when justified.
If one of these is a blocker, email. Priorities follow real demand.
Free to use. Closed source by design. Invitations in waves.
No nurture sequence.