LIT Platform#
The labs make the models smarter. LIT gives your AI a seat on your team, on your machines, with rules and a record. A seat is a credential the team holds and a member uses, so the agent works as a colleague with an identity rather than as a script with a key. Rules are per-team policy: which providers, models, tools and budgets each person's agent may use, where deny always wins and no policy means nothing changes. The record is the channels, transcripts and audit log that belong to you, not to the vendor.
LIT runs any harness and any model, on seats or tokens. It is self-hosted, on your machines or ours through a partner, and that is table stakes rather than the point. The point is the team.
One workplace, three doors#
A door is where you are, not who you are. The receptionist, the project manager, the inside sales rep, the marketing intern and the engineer all use the same workplace; they come in from different places during the day.
In a browser tab. Nothing to install. Sign in with the company login and a seat and a provisioned agent are waiting; one OAuth and you're working. This is the door for most of the company: ask the agent how something works, keep one channel per project, get a summary every morning. IT holds the seats, sets policy and reads the audit log from the same place, but that's a role, not a door. Seats & Policy →
On your own machine. For work that lives in files: drafts in a folder the agent can read, a repo beside Claude Code, local models when the data can't leave the laptop. Docked panels, search across everything the channel has ever said, one person's workplace that connects to the team's when there is one. Desktop →
In your pocket. Native iOS and Android apps, in beta, plus a mobile web UI with voice-first input. Dictate a follow-up between calls, check what the agent did overnight, redirect it, tap an approval card when it asks a question. Mobile →
One record and one policy across all three, because host, team and channel are one address space. Nothing is synced between "versions" of the product; there is one workplace with three ways in.
Team#
Seats & Policy#
Couldn't be trusted by IT → seats, policy, audit, kill switch. An admin holds the subscription seats and API keys for the team; members use them without ever seeing a secret. Policy sets provider and model allow-lists (by name or by family), spend budgets per team, group or member, MCP allow-lists and excluded paths. Deny wins. A team with no policy behaves exactly as it did before, so turning governance on is never a migration.
Channels & the Record#
Couldn't remember → channels with persistent memory. A channel is a named workspace where context compounds for months: messages, the agent's work, pinned decisions, skills, search and a calendar of what happened when. Transcripts and the audit log sit underneath. When someone leaves, their channels and transcripts go to a named successor instead of vanishing with the account.
Identity & Security#
Keycloak single sign-on, federated to the identity provider you already have. Owner, admin, member and group roles. On team deployments, tools execute as the authenticated user over SSH, inheriting that person's real filesystem permissions rather than a service account's.
Agents#
An agent is a prompt, a model, a harness, a set of MCP servers and a set of skills. Admins can provision one as a template so every member gets the same colleague on day one.
Agents → — provisioned agents, dialog cards (the agent asks, you tap), delegation between agents.
Heartbeat → — couldn't work while you slept. The agent wakes on an interval, does its work, posts to a channel and sleeps again. Supervised from a phone.
Scheduled tasks → — the same thing on a calendar.
Self-improving agents → — agents that curate their own notes so the next session starts smarter than the last.
Memory → — file-based, AI-curated, human-readable. Why we didn't build it on RAG.
Any harness, any model#
LIT spawns the vendor's own harness, Claude Code, Codex CLI or Gemini CLI, instead of reimplementing the loop around the model. That is why a Claude Max or Team subscription seat and a pay-per-token API key can sit side by side in the same workplace with the same record: the subscription is used the way the vendor intends, by the vendor's own tool. Direct API access to Anthropic and Google, local models through Ollama, and a model picker, arena and throttle to compare and route. Couldn't be anything but Claude → switch models without losing the record.
Skills, apps & integrations#
Skills are SKILL.md files, personal, team-shared or bundled, that teach an agent a procedure your team has already worked out. Apps are small front-ends that run on the platform against your team's agents. MCP servers connect Google Workspace, Jira and whatever else your team already uses, with per-agent tool configuration and team-level allow-lists.
The record is yours#
Everything runs on your machine or your partner's. Sessions, channels, transcripts, memory and credentials live under ~/.config/lit/ per user; nothing is sent to LIT.AI servers. Switch AI providers and the record stays where it was.
On team deployments, tools run as the authenticated user, with that person's permissions and no one else's, and every action is on the audit log. Proprietary market data or patient records are not a special case; they are the normal case the design assumes.
Getting started#
| Option | Who it's for | Commercial use |
|---|---|---|
| LIT Desktop | One person, same workplace — free for personal use | License → |
| pip install | The single-user server on your own machine, free for personal and non-commercial use | License → |
| Through a partner | Teams. Partners deliver the platform under their brand, administer policy, and own the relationship | ✓ |
| AWS Marketplace | Teams that want a pre-built image in their own AWS account | ✓ |
| Self-hosted | Organizations deploying multi-user on existing infrastructure | ✓ |