An agent platform — meaning the runtime your agents live in. Persistent memory, parallel workers, native tools, scheduling. Bring your own LLM brain or use ours.
ChatGPT forgets you between sessions unless you fight with its custom instructions. Lara remembers what you told her two months ago, the names of your team, your projects, what you tried that failed, what you decided. Every chat builds on the last.
Most AI assistants think one thought at a time. When you ask Lara to research a market, review a contract, or brainstorm a launch, she can send out five copies of herself to work on it in parallel — then read all their answers and give you the summary. What takes a single chatbot 10 minutes takes Lara 30 seconds.
Tell Lara once: "every weekday at 7am, scan my inbox and tell me what's actually important." She'll do it. Forever. You don't have to remember to ask. You walk in, your coffee's brewing, your morning brief is waiting. Most assistants only work when you open a tab.
Drag in a voice memo, she'll transcribe it. Drop in a screenshot, she'll read it. Ask for a 30-second background music loop, she'll make one. Have her clone your voice from a 10-second clip and read your script back. Most chatbots talk. Lara does.
Lara runs on infrastructure you control. Your conversations don't get scraped to train someone else's next model. If you ever want to leave, you take your data with you — every memory, every conversation, every schedule, exportable. Most subscriptions own you. You own this one.
$29/month for the Pro plan. Up to 25,000 messages a month — that's ~833 a day, more than anyone uses. No per-token charges, no surprise bills, no "you've hit the cap, that'll be another $40." Just a number you know going in.
Same Lara, three ways in. She remembers you across all of them — chat in a browser, switch to your terminal, build her into your own app. One persona, one memory.
Open a tab. Sign in. Start talking. The web playground at play.brownforces.io is the simplest way to use Lara — chat with her, see her memories, set her schedules, manage upgrades. Currently invite-only — request access via email.
Coming soon — install once, run kelex in any project directory — chat, edit code, spawn parallel reviewers, attach files with @filename. She remembers each project separately.
HTTP API. Standard auth (tenant ID + bearer token). Streaming responses. Use Lara as the brain behind your own app, automation, or SaaS.
macOS, Linux, Windows (WSL). Requires Python 3.10+. CLI in development — not yet on PyPI. Email to join early access →
# 1. Install pip install kelex-cli # 2. Log in (you'll be asked for tenant_id + api_token from your signup email) kelex login # 3. Run it in any project directory cd ~/code/your-project kelex › hello, lara # Useful commands inside the REPL: › /quota # show your tier + usage › /memory project decisions # search what she remembers › spawn 5 helpers to review @src/auth.py › remind me every monday 9am to triage issues
CLI is open source. Self-host customers point it at their own server with kelex login --base https://your-domain.
A few ways real people are using Lara today.
Uses Lara as a CTO-in-her-pocket. Drafts product specs, reviews emails, brainstorms pricing, summarizes investor calls. Lara remembers every customer interview transcript, so when she's deciding what to build next, she can ask "what did Sarah say about onboarding back in February?"
Each team member has their own Lara, and Lara helps run the routines. Daily standups summarized from Slack. Vendor invoices flagged. Customer issues triaged. The "every Monday morning" routines that nobody has time to do, but suddenly do themselves.
Installs pip install kelex-cli, runs kelex in each project. Lara reads the codebase via @filename, suggests edits as diffs the CLI applies on confirm, runs 5 parallel code reviewers on PRs, and remembers every architectural decision the team has made. Claude Code / Cursor, but with memory that doesn't reset and parallel reasoning built in.
One Lara per client — each remembers that client's context, history, decisions, contracts, players. Switching clients = switching personas. No more "wait, who said that?" Confidential by design (each client's Lara can't see the others).
The actual competitive set. Letta is a toolkit for developers building agents. OpenClaw is a self-hosted agent that's racked up 138 CVEs. Kelex is the agent platform that's already built, multi-tenant, and designed defensively.
| Letta | OpenClaw | Kelex | |
|---|---|---|---|
| Ready-to-use product or build-your-own? | Toolkit — write code, define memory blocks, register tools | Self-host install, configure yourself | Already built. Sign up, get tenant, agents work today. |
| Multi-tenant from day one | Cloud version only | Single-user installs — each one is its own attack surface | Tenant-keyed schema, FK cascade isolation |
| Native parallel workers (swarm) | Build it yourself | No | delegate(N) primitive, up to 16 workers + synthesizer |
| Self-scheduling agents (cron) | No | No | schedule.create, server-side cron |
| Multi-brain routing (BYOK) | Per-agent provider config | One brain at a time | brain.ask(provider) — Claude, GPT, Kimi, DeepSeek, local — per persona or per call |
| Multi-modal tools (voice / vision / music / video) | Text only | Text only | 8 tools built in |
| Security posture | Open source, reasonable defaults | 138 CVEs · ClawBleed RCE · banned by China | Hashed tokens · behind tunnel · quotas · no shell access · ~3K LOC auditable |
| Includes a terminal CLI | Letta Code (separate product) | No | pip install kelex-cli (coming soon) |
OpenClaw is the most-starred AI agent on GitHub — and a security disaster. The reason we built Lara is simple: we wanted what OpenClaw promised, without the part where running it on real data gets you breached.
Same self-hosted agent vision. Different design choices, made with security as a constraint not a retrofit.
| Risk | OpenClaw | Lara on Kelex |
|---|---|---|
| Server exposed to public internet | Default Docker opens ports. ~135K instances scanable. | Behind Cloudflare Tunnel — origins aren't publicly addressable. Zero open ports needed. |
| Auth token theft (the ClawBleed bug) | Tokens stealable in plaintext via API exploit. Active in the wild. | Tokens stored as sha256 hashes only. We literally cannot leak plaintexts because we don't have them. |
| One bug = many tenants breached | Each single-user install is its own attack surface. | Multi-tenant from day one. tenant_id on every row with cascading FKs. Cross-tenant requires breaking auth middleware. |
| Runaway loops burning compute / money | "Excessive energy use" cited in the China ban. | Tier-aware quotas enforced server-side. Hits 25K msg/mo? Hard 429. No runaway bills. |
| Tools with shell / file access | Open by design — agents can run arbitrary code. | Tools are HTTP calls to constrained endpoints with JSON schemas. No shell. No file write unless you wire it. |
| Webhook spoofing | Custom signing, often missing. | HMAC signature verification + idempotency log. Replay attacks impossible. |
| Service running as root | Default Docker setups commonly run as root. | Runs as a regular user. No privilege escalation if compromised. |
| Attack surface size | Sprawling. 138 CVEs implies many vulnerable code paths. | Single FastAPI app + Postgres. ~3,000 lines. Auditable in an afternoon. |
If you tried OpenClaw and the CVE count gave you pause — good instinct.
That's exactly why we built Lara.
Infrastructure pricing — we provide the agent runtime, you bring the brain (or use ours). No per-message billing on local. External brain calls go directly to your provider account.
If you're not sure whether this is for you.
kelex-cli, but she's also your assistant for ops, customer research, scheduling, multi-modal work (voice/vision/music/video). She remembers across every conversation. She wakes herself up. And the same Lara runs on the web, in the terminal, and via API — all sharing one memory. Aider/Cline are wrenches. Lara is a personal employee.pip install kelex-cli will get you the kelex command — a chat REPL in your terminal that remembers each project separately, can read your files (@filename), spawn parallel helpers, and apply code edits. Same Lara, same memory as the web — she remembers you across both surfaces. Mac, Linux, Windows (WSL). See the install section above.I'm Salvador, the founder of Brown Forces Technology Studio. I'd been paying for ChatGPT Plus and Claude Pro and hitting the same walls — they don't remember me, can't run errands while I sleep, can't fan out work in parallel.
OpenClaw promised exactly what I wanted: a self-hosted agent with memory that worked. So I went to install it. Then I read the security write-ups. 138 CVEs in 6 months. ClawBleed actively exploited in the wild. China banned it from state agencies for data leaks. I wasn't going to put that on my server next to real customer data. The risk wasn't worth it.
So I built what I actually wanted — same vision, defensive design from day one. Multi-tenant, hashed tokens, hard quotas, no shell access, behind a tunnel. I use Lara every day. Now anyone can.
Email [email protected] with the tier you want. You'll have a Lara of your own within a day.