One agent. A whole organisation. Engineered for the long run.
Meerkat is an open-source platform for building and operating agents. Run one from a command, embed a team in your product, or keep hundreds working across an organisation for months. Agents retain their identities and history as work changes and processes restart. Execution, coordination, memory, integrations and security share foundations built for native asynchronous operation, with critical lifecycles model-checked. Mix cloud and self-hosted models in the same team, on infrastructure you control.
Codex, Claude Code, Cursor, Copilot and Pi can turn an ambitious brief into working software, tackling complex problems for days with little intervention. They are built around delivering a result, with a person setting direction and judging the outcome.
Meerkat is built for agents that become part of a running operation. They respond to events, coordinate work and carry responsibilities for months. Tasks come and go; the agents keep their identities, memory and roles.
Teams can span machines; agents communicate directly and coordinate shared work. Your application controls membership and connections. Critical lifecycle rules are model-checked, while scheduling, recovery and access control govern how work continues and under whose authority.
The same foundation is composable: compile in only what you need for an ESP32 or WebAssembly, or use the full integrated platform, including Meta, Connect and Recall. Open source and provider agnostic throughout. That is what we mean by an operational agent platform.
A complete operational platform, composable down to the runtime. Compile in the capabilities you need: a minimal, low-footprint agent on an embedded device, or a full-featured harness running in a data center. At its fullest, Meerkat integrates agent teams, memory, integrations, identity, security and operations with Recall, a complete knowledge system for evidence, relationships and traceable conclusions.
Give an agent access to your calendar without permission to delete it. Let it prepare an invitation, but ask you before sending. Meerkat separates who the agent acts for, what it may do and which actions need approval, with permission checked when the action is about to happen.
Acting on behalf of Maya, within a calendar-specific grant.
The requester, executing agent, represented person and connected account have distinct identities. Grants can narrow actions, resources, recipients, lifetime and delegation depth. A helper cannot widen its parent's grant; expiry and revocation are checked along the chain.
Meta's ABAC rules match people and groups against agent roles and labels, with separate controls for conversations, memory, topology, work and administration. Native operation grants and each resource's own policy govern execution. Seeing an agent in the console does not grant its tools.
Policy can require local permission alone, a bounded model review, or fresh human consent. Reviews bind to the exact proposed action. Changed arguments, accounts or recipients need a fresh check. Neither a reviewer nor an approval can override missing permission.
Required sandbox profiles constrain filesystem access, network connections, environment and inherited handles for confined shell and MCP processes. Linux and macOS have native backends. If the host cannot enforce the selected requirements, the launch is refused.
Delayed actions are checked again before execution. A previous allow is not a permanent permit. A refused operation gives the agent feedback to choose another permitted action, while independent allowed work can continue.
Audit records distinguish permission decisions, reviews, consent consumption, execution attempts and returned outcomes. They stay associated with the work that caused them. Uncertain outcomes remain explicit, so an approval or a missing result is never mistaken for a completed action.
Configure governance at the runtime, resource and process boundaries your application uses. Console access rules and provider sign-in do not activate those controls on their own. Information already read into a model context still requires appropriate handling rules and separation.
Meerkat ConnectPreview
Describe the job, such as turning a week's Slack updates into an email draft and a changelog. The agent identifies the capabilities it needs, finds integrations in the library or researches beyond it, and prepares the setup. Your application connects the tools with the accounts and access you choose.
100+integrations in the library
A selection. Supported actions and setup vary by integration.
“Every Monday, summarise last week's #releases in Slack, draft an email to the team and add it to our changelog.”
Start with the job Describe the outcome. Integrations don't need to be named up front.
Identify the capabilities Reading a Slack channel and replying in Slack as the agent are separate capabilities. So are drafting and sending.
Library first, then beyond Two needs match the library. For the third, the agent researches packages and picks one.
Setup, prepared Setup code validates the package choice. The plan lists the accounts and access each tool needs.
Your application connects You choose accounts and access. Your application authorizes the install, holds the credentials and attaches the tools.
Services the agent reads from and works in, like the Slack, Gmail and changelog tools above.
Use Slack, Telegram or another messaging app as the interface to your agents. Set up the conversation channel separately from the tools agents use to read data or take actions.
Meta brings agents together as an application, with memory, routing, access control and approvals. Its console is a ready-to-use interface and a UI framework you can adapt: change the layout and branding, add application-specific views, or build your own interface from reusable React components.
The system view: the agent topology, an initiative agent's generated progress card, and a person's own agent tracking shared work.
The topology view: review workers cluster around the work they were sent to do, with their links to the rest of the system.
Every agent with its role, profile, state and lease. Select one to see its session and the agents it is wired to.
The same system as a tree of roles: personal and initiative agents, the analysis agents, and the workers they send out.
A person asks their own agent about an initiative. It updates the shared work in the WorkGraph and records the confirmation as evidence.
A shared knowledge base for agents. Recall connects documents, people, projects and their relationships to the claims they support. When two records disagree, it keeps the evidence separate from the current conclusion, so agents can trace an answer to its source and reconsider it as new information arrives.
Meta memory carries lessons, decisions and procedures forward for an agent. Recall connects shared facts, relationships and evidence so agents can reach and revisit conclusions.
What's blocking Friday's release?
The agent's accumulated experience
Lead with blockers.
Check approvals before saying ready.
Guides how it approaches the task
Shared knowledge, linked to evidence
Grounds its answer in current evidence
Security review is still open. The build passed.
Meta can share reviewed memories with a team, too. Both systems keep provenance and enforce configured access. Neither requires retraining the underlying model.
Both systems are in daily use on Core and Meta. One coordinates a household from a small computer at home. The other follows work across an engineering organisation. Their agents respond to schedules, incoming information and each other between conversations with people.
Each person has an agent, reached by chat, console or voice. Specialists handle the calendar, school, home, cameras, network, finance, health and research. Triage reads incoming mail, app updates and documents, then files or routes what matters. Specialists return findings to the person's agent, which decides what they need to hear. Agents keep their own experience; a shared knowledge base connects household records and their sources.
Agents follow the code, tickets and plans behind each initiative, checking reported progress against the underlying work. Each repository has continuing domain agents that consult across repositories. Each initiative has an agent tracking its goals, work items and evidence in a WorkGraph. Analysis agents send out workers for reviews and investigations. People hear from their own agents, and a reporting agent prepares the weekly summaries.
Use the command-line tool, embed the Rust library, or call Meerkat from Python or TypeScript. The SDKs fetch the server they need. REST, JSON-RPC, MCP and browser interfaces reach the same runtime.
Drop in a server alert and watch ten agents coordinate. The harness runs locally in your browser as WebAssembly, with model calls over an API.
Enter the office# install (or: cargo install rkat) $ brew install lukacf/meerkat/rkat $ export RKAT_OPENAI_API_KEY="sk-..." # one agent, one job $ cat meeting-notes.txt | rkat run \ "Turn these notes into dated to-dos." # continue the same session later $ rkat run --resume last \ "Which of these are due this week?"
# Cargo.toml [dependencies] meerkat = "=0.9.0" tokio = { version = "1", features = ["full"] } // in your service let mut agent = AgentBuilder::new() .model("claude-opus-5-5") .system_prompt("You are an incident triage component.") .build(llm, tools, store) .await?; let result = agent.run(raw_alert_text.into()).await?; println!("{}", result.text);
$ pip install meerkat-sdk==0.9.0 $ export RKAT_OPENAI_API_KEY="sk-..." # main.py import asyncio from meerkat import MeerkatClient async def main(): async with MeerkatClient() as client: session = await client.create_session( "What is the capital of France? Answer in one sentence." ) print(session.text) asyncio.run(main())
$ npm install @rkat/sdk@0.9.0 $ export RKAT_OPENAI_API_KEY="sk-..." // main.ts import { MeerkatClient } from "@rkat/sdk"; const client = new MeerkatClient(); await client.connect(); const session = await client.createSession( "What is the capital of France? Answer in one sentence.", ); console.log(session.text); await client.close();
# paste into your coding agent
Install Meerkat with: brew install lukacf/meerkat/rkat Use my provider key from RKAT_OPENAI_API_KEY, ANTHROPIC_API_KEY or GEMINI_API_KEY. Read https://docs.rkat.ai/llms.txt before you write any code. Then embed Meerkat in this project through the Python SDK (meerkat-sdk), the TypeScript SDK (@rkat/sdk) or the Rust crate (meerkat), whichever fits it.
Meerkat
