Skip to main content

Install

The tap supports macOS and Linux and installs rkat, rkat-rpc, rkat-rest and rkat-mcp. Without Homebrew, use the release archives (see Linux without Homebrew).

Linux without Homebrew

Each GitHub release has one archive per binary per target, plus checksums.sha256. Download the binaries you need, verify them, and put them on your PATH:
Add rkat-rest or rkat-mcp to the loops if you need them. rkat-rpc is the binary the Python and TypeScript SDKs drive. The Linux binaries need glibc 2.31 or later; musl distributions such as Alpine are not supported (see Linux glibc floor).

Configure One Provider

The fastest path is a dedicated Meerkat environment variable:
Provider-native names also work. For OAuth, Azure, cloud IAM, managed-store credentials, and tenant-scoped bindings, see Auth.
Meerkat’s OpenAI/global catalog default is gpt-6-astra. Explicit model and binding defaults are preserved. Catalog support does not guarantee API access or support through Azure, Copilot, or custom backends. Select an available model explicitly or configure a binding default_model when needed.

Run A First Session

With the OpenAI variable above:
With another provider, select a matching model:
The SDKs connect to rkat-rpc; the CLI constructs the same runtime-backed session in-process.

Share State Across Surfaces

CLI commands derive a stable workspace realm by default. A bare RPC/REST/MCP server or SDK connection creates a fresh isolated realm. Pass one explicit ID when several processes should see the same sessions and config:
The processes must also resolve the same physical storage root/provider. See Realms.

Add An MCP Server

Config changes apply to newly built/resumed agents. MCP servers normally connect in the background and tools appear as each server becomes ready; --wait-for-mcp makes the first prompt wait for configured servers. For a remote streamable-HTTP server that signs in with OAuth (replace the example URL with the service you use). discover binds the account the service verifies at the first login; use --oauth-account-selection unverified only for a server that cannot prove an account:

Bound One Run

  • --max-tokens caps LLM output per turn.
  • --max-tool-calls bounds tool use for the run.
  • --max-duration bounds the run’s agent lifetime; a deadline is a hard terminal failure rather than a warning.
For aggregate turn deadlines, lifetime token budgets, retry policy, and admission limits, use realm configuration.

Add Built-In Capabilities

The full preset enables the product CLI’s broad tool composition, including shell, MCP, skills, comms, schedules, WorkGraph, and mob orchestration where compiled:
Use --allow-tool and --block-tool for a narrower first-turn scope. Shell access and other effectful tools should be granted deliberately.

Next Steps