PurrCat
Star

Build an agent that works your way.

Let expression no longer be a lossy compression. PurrCat treats intent as the thing worth preserving — how it enters the system, how it flows through execution, and what is worth keeping afterwards.

Install PurrCat Read the docs
MCP servers Agent Skills Docker sandbox Electron desktop DAG workflows

Ready. What are we making today?

Help me cut the demo video: grab the three clips in ~/footage/demo, drop the boring middle, caption it from the audio, pick a music bed that fits the pacing, keep it under 60 seconds, export 9:16 for shorts and 16:9 for the site, use the $ffmpeg skill and the $caption skill, follow the $brand-edit workflow my team wrote, and pull the brand colours from the asset MCP server. Last time you cut the ending too early.

Ready. What are we making today?

Cut the demo video for me.

  • Autoskillsffmpeg-cut · auto-caption
  • Autoworkflowbrand-edit v3
  • Autoclips3 · ~/footage/demo
  • Autotarget< 60s · 9:16 + 16:9
  • Autoassetsasset-library MCP
  • Automemorybrand colours · last cut ran short

What's inside

Five commitments, one foundation.

From how information enters the agent, to how tasks execute, how memory is organised, and how capabilities expand — you orchestrate all of it the way you work.

00

Information-Driven Agent Design

The model decides capability; information decides direction.

  • UI plugin architecture. Fold your working context into UI panels — a click can communicate what a paragraph used to.
  • Orchestrated information flow. Describe the agent's execution as a configurable DAG and orchestrate the loop with a single yaml file; you decide when information gets injected.
  • Multi-layer hybrid memory. Short-term working memory, MEMORY.md, and PurrMemo's episodic, semantic, and knowledge-graph layers, injected in stages.
  • Sub-agents that share context. Their own branch id, model instance, and event loop — over one shared contextual foundation.

01

Everything Is Plugins

Beyond the kernel — capabilities, flows, and interfaces are all extensible.

  • Tools, skills, sensors, agent loops, workflows and UI panels all plug in or get replaced.
  • One yaml file configures information injection, tool-use checks, and loop-exit conditions.
  • Six extension types: skill, ui-plugin, mcp, agentloop, sensor, graph.

02

RSI Self-Evolution System

Not only using capabilities, but building, evaluating, and iterating them.

  • The agent builds its own plugins: Trace2Skill, MCP tool building, sensor building.
  • Generated skills ship with a SKILL.md and an evals.json set, and must pass the built-in evaluator.
  • Hot reload, standalone HTML evaluation reports, and an iteration archive that can roll back.

03

Secure Self-Hosting

The agent has its own workspace; your host machine is still yours to control.

  • Commands and files live in a dedicated Docker container with persistent workspace and logs.
  • Imports are size- and path-validated; exports can trigger a Git snapshot for audit.
  • GOAL.md plus a heartbeat cycle lets the agent patrol and report while idle.

04

Efficient Context & Tool Management

Reduce unnecessary tool information; keep a stable long-session cache.

  • Key manager binds a session to one key, keeping the long-session prefix cache intact.
  • Tools and skills are stored then retrieved on demand — an attention problem turned into a retrieval problem.
  • Workflows pass node results as structured data, cutting repeated restatement between agents.

See it run

You used to type the brief. Now you press one button.

Three everyday jobs, one plugin surface. Pick a job and the panel mounts, the loop switches, and the agent fills in everything you used to type yourself.

AI + video

Build your own workbench

Panel container slot · none

Pick a job above — the panel mounts right here.

Session Loop · Default

Pick a job, then watch the panel press Start.

    Compound

    It gets smarter the longer you use it.

    Skills that are blind-tested until they hold, a memory that learns how you work — and not only those two. Every plugin you switch on evolves the same way.

    Skill factory Sandboxed blind eval

    quant-strategy · iteration trail

    • Iter 133%
    • Iter 267%
    • Iter 3100%
    • build_alpha_pipeline9 / 9 assertions 812s · 684k tok
    • diagnose_leakage7 / 7 assertions 665s · 513k tok
    • decider_usage6 / 6 assertions 431s · 299k tok
    Pass rate3 / 3 · 100%
    Avg time636.2s
    Avg tokens498,549
    Tokens saved−32% · 486k
    Memory Three layers, one recall
    • 01 Working memorySession context, held while the task runs.
    • 02 MEMORY.mdDurable notes the agent writes and re-reads.
    • 03 PurrMemoHybrid recall — events, experiences, a graph.
    eventstop 5 · time-filtered
    experiencestop 5 · vector
    graph nodestop 3 · depth 2

    Quickstart

    Install PurrCat

    Three steps from a clean machine: install PurrCat, launch the desktop app, and make it yours.

    1 Install PurrCat
    Install via

    Run in your terminal, then follow the prompts.

    curl -fsSL https://raw.githubusercontent.com/PurrPod/purrcat/main/install.sh | bash
    What the script does

    It checks for and installs the required prerequisites, clones the source to ~/purrcat, installs the project dependencies, and registers a global purrcat command.

    Windows: launch Docker Desktop once after installation to complete setup and accept its agreement. The Windows flow is well tested; macOS and Linux have more limited coverage.

    2 Launch the desktop app

    Start the Electron desktop app, then confirm the API is listening.

    01 Start the desktop app

    purrcat desktop start

    02 Pin the API port

    The API uses port 8000 by default and quietly moves to the next free port when it is taken. To pin one instead:

    PURRCAT_API_PORT=8123 purrcat desktop start
    3 Make it yours

    Tell it your workflow, and it writes you a workbench of your own.

    • Keep it current.

      Pull the latest source and refresh dependencies in one command.

      Update PurrCat
    • Read how it works.

      Plugins, memory, workflows, and the sandbox, in detail.

      Open the docs
    • Follow along.

      Star the repo to keep PurrCat close and support the project.

      Star on GitHub
    Start building

    Design goal

    Let the agent be not just stronger, but more aligned with how you work.

    Install PurrCat View on GitHub