makakoo
— open project · MIT forever · built by contributors from everywhere
makakoo
— open project · MIT forever · built by contributors from everywhere
╔═══════════════════════════════════════════════════════╗ ║ ║ ║ ███╗ ███╗ █████╗ ██╗ ██╗ █████╗ ██╗ ██╗ ██████╗ ██████╗ ║ ║ ████╗ ████║██╔══██╗██║ ██╔╝██╔══██╗██║ ██╔╝██╔═══██╗██╔═══██╗ ║ ║ ██╔████╔██║███████║█████╔╝ ███████║█████╔╝ ██║ ██║██║ ██║ ║ ║ ██║╚██╔╝██║██╔══██║██╔═██╗ ██╔══██║██╔═██╗ ██║ ██║██║ ██║ ║ ║ ██║ ╚═╝ ██║██║ ██║██║ ██╗██║ ██║██║ ██╗╚██████╔╝╚██████╔╝ ║ ║ ╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═╝ ╚═════╝ ╚═════╝ ║ ║ ║ ║ O S ║ ║ ║ ╚═══════════════════════════════════════════════════════╝
Install it once. Then use it where you already talk to AI: terminal, IDE, Pi, Telegram, Discord, or your own connected agent.
"what do I owe HoCa, and what did we promise in the last proposal?"
"continue the website sprint, check the live page, and publish fixes"
"read these papers and build a careful research brief from my notes"
"draft a reply to this client, but do not send it until I approve"
Commands are for setup and maintenance. Daily work can happen in an IDE chat, terminal chat, Pi, Telegram, Discord, HarveyChat, or an external agent connected through Lope, MCP, subprocess, or HTTP. Makakoo is the home they all come back to.
Start with Pi, Telegram, or Discord plus Makakoo. Best if you want an assistant that remembers without living in a terminal.
Use Warp plus OpenCode or Claude Code plus Makakoo. Best if you want code, docs, releases, and project memory in one workflow.
Add Antigravity/agy, Codex, Qwen, SwitchAILocal, and Lope. Best when one agent should build and others should challenge it.
Pi, Telegram, Discord, and HarveyChat are for people who do not want to live in a terminal. Cortex gives those chats local memory, so the assistant can remember useful things later.
Warp, OpenCode, Claude Code, Codex, Antigravity/agy, Cursor, Vibe, and Qwen are the builder tools. Use them for code, docs, research, releases, and operations.
One local address for models. Agents can use CLI subscriptions, local models, cloud APIs, or automatic routing without every tool learning every provider.
The Brain keeps journals, pages, project notes, and multimodal knowledge. Cortex remembers chat details and can link the same person across Telegram, Discord, and future channels.
Add other agents by command, subprocess, or HTTP. One builds, one critiques, another votes. You get disagreement before you trust the result.
Small background workers and friendly creatures watch logs, docs, inboxes, memory, health checks, and project cleanup while you sleep.
Makakoo works wherever an AI helper can talk: terminal CLIs, IDE agents, Pi, Telegram, Discord, HarveyChat, and connected external agents. Commands are useful, but they are not the product. The product is plain language with memory. Type what you need. Harvey reads the Brain, uses tools, calls Lope when needed, checks the real state, and writes down what happened.
"draft the HoCa follow-up from the proposal notes and keep it human"
uses project memory, preserves commitments, drafts without sending"compare these papers and tell me what claims are actually supported"
extracts claims, checks methodology, writes a reusable brief"create a launch sprint, ask Lope for critique, then execute it"
negotiates scope, runs phases, validates outcomes"what happened with the last Netlify deploy?"
searches journals, reports deploy id, commit, live verification"look at this screenshot, explain the problem, remember it"
uses multimodal tools, writes durable memory when asked"the installer fails on Windows, find root cause and fix it"
checks scripts, release assets, CI, and patches only the broken pathMost AI tools are goldfish. Every session starts from zero. Every context paste is manual. Makakoo is different - it is an operating system for AI agents doing real work: research, writing, client ops, inbox triage, project memory, media analysis, business planning, and software. Persistent memory, multi-model validation, and a growing family of diagnostic creatures keep the house alive.
Claude can remember a client decision. Antigravity/agy can pick up the research trail. Codex can check release state. OpenCode can continue the plan. Pi and HarveyChat can remember conversations through Cortex. They all read the same local Brain, so you stop copy-pasting context between tools.
Lope gives Makakoo a validator ensemble. One agent drafts, others review independently, objective gates run, then a consensus verdict decides what moves. No single-model blindspot. This is how Makakoo handles structural changes without vibes-only trust.
SANCHO can run scheduled background tasks for wiki linting, memory consolidation, index rebuilds, error classification, mascot patrols, Brain sync, and health checks. You sleep. Makakoo watches. Results are logged to your Brain journal.
makakoo infect connects local AI hosts to the same home. Pi, Telegram, Discord, HarveyChat, Lope providers, MCP tools, subprocesses, and HTTP adapters can connect too. You do not need to memorize commands. Ask naturally: "continue the sprint", "what did we decide?", "review this plan", "fix the bug". The body changes. The Brain stays.
Errors can be caught, classified, and clustered by the GYM. The hypothesizer proposes fixes; Lope's validator ensemble reviews them independently; approved fixes queue for human review. Makakoo is built to get better while you sleep - without silently changing your machine.
Local CLI bodies are only the start. Makakoo already has chat channels, Cortex aliases, Lope subprocess and HTTP providers, MCP tools, and the HarveyChat gateway. External agents can join without pretending everything is a terminal.
Makakoo is the house for Harvey, not a code-only plugin. The same memory and tool surface works when you are writing proposals, researching papers, managing clients, checking email, planning a launch, reviewing finances, creating content, or debugging software. The point is continuity: the agent should know what happened before and act from evidence.
"check the latest client emails, draft replies in my voice, keep the thread open, and never send without approval."
"turn my rough notes into a clear proposal, keep pricing consistent, and save the polished version where the project expects it."
"read these sources, extract claims, compare methods, and remember the important findings for the next session."
"summarize this video, listen to the voice note, inspect the screenshot, then ingest it if I say remember it."
"check what is actually shipped, then draft posts, website copy, and launch arguments based on live proof."
"create a sprint, ask the Lope team for critique, choose the safest route, execute, audit, and log the result."
"update my CV from the latest targeted draft, evaluate opportunities, write recruiter replies, and track follow-ups."
"fix the bug, verify CI, publish the release, check production, and tell me exactly what is done versus still risky."
There is a biological idea called host manipulation: a fungus enters a host and changes what the host does. The famous example is Ophiocordyceps unilateralis, the zombie-ant fungus. Creepy in nature. Useful as metaphor. makakoo is the friendly, consent-based version for AI hosts: CLIs, IDE agents, chat apps, and connected external agents.
You run makakoo infect on purpose. It writes a bootstrap into each agentic host so the host stops behaving like an isolated goldfish and starts behaving like a body connected to the same local mind.
Claude Code, Codex, OpenCode, Antigravity/agy, Cursor, Vibe, Qwen, Pi, Telegram, Discord, or any future agentic host. Each body has different strengths.
The bootstrap block. Small, readable, reversible. It tells the host where the Brain is, what persona to use, what tools exist, and what rules matter.
The shared Brain on disk: journals, pages, graph search, vector recall, multimodal knowledge, auto-memory, and durable project context.
The host now answers from memory, asks before sending, uses tools, logs important work, calls Lope for critique, and resumes where another body stopped.
Nature's parasite forces an ant into a death grip. makakoo gives an AI body a life grip: memory, continuity, safety rails, tools, and a family. You stay in control. The host gets smarter. The Brain gets stronger.
Makakoo is not just a memory folder and not just a CLI installer. It is the local house Harvey lives in: brain, bodies, skills, tools, safety rails, background workers, mascots, multimodal recall, and Lope's adversarial review loop. Everything is file-backed, inspectable, and designed to survive context resets.
Logseq-style journals, wiki pages, FTS search, graph links, vector recall, and multimodal knowledge for PDFs, images, audio, and video.
makakoo infect wires Claude Code, OpenCode, Antigravity/agy, Codex, Cursor, Vibe, Qwen, and future hosts into the same persona and memory. Pi, Telegram, Discord, and external agents can connect too.
One local OpenAI-compatible power socket for agents. Route to CLI subscriptions, local models, cloud models, or automatic provider selection without every tool learning every API.
Shared MCP surface: browser control, Google Workspace, Telegram bridge, image/audio/video understanding, write grants, skills, agents, docs MCP, local shell work, and local S3/Garage state.
Long-term chat memory for HarveyChat. Telegram and Discord messages can resolve to the same person through explicit aliases, then recall useful memories before the next answer.
Permission grants, explicit outbound-message approval, sandbox-aware writes, evidence gates, CI checks, release hashes, and no telemetry by default.
The validator lion in the mycelium. Multi-CLI reviews, sprint negotiation, execution gates, scorecards, SARIF, votes, comparisons, memory, and council deliberation.
Olibia, Pixel, Cinder, Ziggy, and Glimmer give the system a visible operating layer: patrols, diagnostics, logs, doc audits, and garden cleanup.
Every mascot has a name, a species, stats, a personality, and one job. They are free-form: users can create their own creatures and attach them to whatever they care about - inbox patrols, proposal checks, research digestion, budget reviews, media indexing, project hygiene, or code health. They run on a schedule, write flavor lines, and make invisible automation feel alive.
Species: Misttooth · Stats: DEBUGGING 53, SNARK 17
Tails daemon logs for fresh error bursts. Traceback, 401 Unauthorized, Error: - if it's in the logs, Pixel sees it. Every 2 hours.
"Pixel hums ominously: no error music today."
Species: Stormnub · Stats: DEBUGGING 65, CHAOS 22
Byte-compiles every Python plugin entrypoint changed in the last 24 hours. Catches broken entrypoints before the next SANCHO tick crashes on them. Every 4 hours.
"Cinder purrs - changed entrypoints compile clean."
Species: Blazebit · Stats: SNARK 60, WISDOM 19
Lints SKILL.md files for completeness. Reports structural gaps and tracks coverage trends over time. Daily.
"Ziggy sighs: several skills need sharper instructions."
Species: Lunanub · Stats: PATIENCE 56, WISDOM 16
Patrols Brain pages for TODO, FIXME, XXX markers in files older than 14 days. Archives stale lead pages. Keeps the garden tidy. Daily at 22:00.
"Glimmer trims gently - garden is orderly tonight."
Patrols client threads, flags money opportunities, drafts replies, never sends without approval.
Reads papers, extracts claims, tracks open questions, reminds you what still needs evidence.
Watches invoices, subscriptions, trading logs, budgets, or revenue leads.
Checks brand tone, assets, launch posts, scripts, or design consistency.
A mascot is just a named patrol with memory, personality, schedule, and a job. Pick a species. Pick a voice. Attach it to a folder, workflow, inbox, project, or goal.
Most AI tools forget everything the moment you close the tab. Makakoo doesn't. The brain is a multi-layer, multi-format memory system that every CLI reads and writes simultaneously. Plain markdown on disk. Obsidian-compatible. Logseq-native. No lock-in. No cloud. Your data stays on your machine forever.
Every significant decision, bug fix, and discovery can be logged. Searchable by date. Readable in any editor. Your AI gets a durable diary instead of starting from zero every morning.
~/makakoo/data/Brain/journals/2026_04_27.md
Every person, project, company, and technology can get a wiki page with bidirectional wikilinks. Like a personal Wikipedia maintained by your agents.
~/makakoo/data/Brain/pages/Makakoo OS.md
Every [[wikilink]] is a relationship. "Harvey built Makakoo." "Makakoo uses Logseq." The graph learns who did what, what depends on what, and what matters most. FTS5 full-text search + vector embeddings + LLM synthesis.
SQLite FTS5 + vector search + entity graph with PPR ranking
After significant work, agents can write structured memory entries. No manual note-taking ritual. The brain grows through local files that every connected body can read on its next turn.
~/makakoo/data/auto-memory/
Index videos, audio files, PDFs, and images into the knowledge store. Qdrant-backed multimodal indexing with model-provided embeddings. Search what you ingested - not just text. Every connected agent can recall it.
Qdrant + embedding model + multimodal chunks
~/makakoo/data/Brain/ and browse your AI's entire memory as a graph of linked notes. No export. No sync tool. No proprietary format. It's just markdown. It's yours.
Makakoo is the brain. Every connected AI surface is a body: terminal CLI, IDE chat, Pi, Telegram, Discord, Claude, OpenCode, Antigravity/agy, Codex, Qwen, or an external agent. You talk in plain English, not command syntax. They share journals, wiki pages, Cortex chat memory, graph search, model routing, and MCP tools. Switch bodies like switching tabs. The Brain stays put.
~/makakoo/data/Brain/. Journals. Wiki pages. Knowledge graph. Vector search. All markdown. All yours.makakoo infect writes the bootstrap into local hosts; Pi, Telegram, Discord, HarveyChat, Lope providers, MCP tools, subprocesses, and HTTP adapters connect more bodies.~/makakoo/data/auto-memory/. When one CLI writes a durable note, the next CLI can read it from the same local source. No context-paste ritual.SANCHO is the daemon layer for Makakoo. It schedules wiki linting, memory consolidation, index rebuilds, error classification, mascot patrols, Superbrain vector sync, and briefings. Tasks write results to your Brain journal, so the operating history stays visible.
wiki_lint scans pages for broken links. memory_consolidation clusters daily facts. index_rebuild refreshes the search engine. superbrain_sync_embed catches orphaned vectors. dream runs memory consolidation cycles. daily_briefing summarizes your journal. All local and inspectable.
every task logged to data/Brain/journals/YYYY_MM_DD.md
Pixel patrols daemon logs for error bursts. Cinder compiles changed Python entrypoints. Ziggy audits SKILL.md files for gaps. Glimmer prunes stale TODOs from Brain pages. But these are just the starter creatures. Users can hatch new mascots for anything: clients, papers, budgets, health checks, creative projects, or house rules.
data/mascots/{pixel,cinder,ziggy,glimmer}/
Layer 1: Error funnel captures every failure to data/errors/. Layer 2: Classifier labels them (environmental/skill/code/user). Layer 3: Hypothesizer generates fixes, scores baseline vs improved. Layer 4: Lope validator ensemble reviews independently - PASS moves to approved, anything else moves to rejected. Layer 5: Human reviews the approved queue.
errors → classify → hypothesize → lope gate → approve
On schedule: checks gateway health, AI provider capability, HarveyChat status, plugin entrypoint health, key process liveness, and SANCHO task results. Returns aggregated healthy/unhealthy verdict. Logged to your journal.
makakoo_heartbeat - 60 min interval, 24/7 coverage
Errors, bad answers, failed tools, broken commands, and user corrections are saved as evidence.
The GYM labels what happened: environment, skill gap, code bug, stale docs, model mistake, or user workflow mismatch.
Harvey proposes fixes and tests: better prompt rule, patch, new skill, docs update, or safer workflow.
Lope and objective gates review the fix. Bad ideas get rejected. Good ones wait for approval.
Accepted lessons go into the Brain and auto-memory so every future CLI body improves.
Install commands set up the house. After that, you do not operate Makakoo by memorizing commands. Talk wherever you already talk to AI: Claude Code, Codex, OpenCode, Antigravity/agy, Cursor, Vibe, Qwen, Pi, Telegram, Discord, or HarveyChat. Makakoo is the layer underneath that makes the chat durable.
"check my inbox for real opportunities and draft the replies"
agent filters noise, uses context, drafts only, waits for approval"turn these rough bullets into a client proposal and PDF"
Makakoo keeps client facts, pricing, and prior decisions consistent"what did we decide about Windows install support?"
Makakoo searches journals, pages, and memory instead of guessing"look at this screenshot and remember the bug for later"
multimodal tools describe media, ingest when asked, and make it searchable"write the launch post, but check the live repo truth first"
agent verifies release state, then drafts from evidence"plan my week from the calendar, inbox, and project deadlines"
agent combines workspace context with calendar and remembered priorities"summarize this podcast and add the useful parts to my knowledge base"
audio/video understanding becomes future-searchable knowledge"use Antigravity as critic, Codex as implementer, and Harvey as boss"
Makakoo coordinates agent roles while the Brain keeps continuityLope is a separate MIT-licensed plugin and CLI harness, but inside Makakoo it is family. You can run Lope commands directly, or just ask in normal chat: "ask the Lope team", "get a vote", "review this file", "audit the sprint". Makakoo gives every agent a shared brain; Lope makes those agents disagree productively before work ships.
🦁Lope asks other AI bodies for independent opinions: Claude Code, OpenCode, Antigravity/agy, Codex, Vibe, Aider, Ollama, pi, Qwen, Copilot, or any custom subprocess or OpenAI-compatible HTTP provider. Use it for engineering, business, research, writing, operations, marketing, launch planning, and hard decisions. Not just code.
$ lope negotiate "launch Makakoo public beta" --domain business $ lope execute SPRINT-launch.md $ lope audit SPRINT-launch.md
"ask Lope to critique this launch plan before I post it"
Validators check positioning, proof, risks, and missing audience assumptions."have Lope review this research protocol for weak methodology"
Validators challenge evidence, sampling, claims, and validation criteria."compare these two proposal versions and choose the clearer one"
Lope can compare artifacts against criteria, not just source files."review this diff for release blockers, then tell Harvey what to fix"
Lope fans out a technical review and turns critique into action.Draft sprint docs through multi-round validator review for engineering, business, or research. Validators push back on scope creep, missing criteria, and unverified assumptions until consensus or escalation.
Run phases one by one. Each phase gets two-stage review: spec compliance, then quality. NEEDS_FIX retries with concrete instructions. PASS advances.
Generate the scorecard: phase verdicts, confidence, duration, overall status, and a journal trail you can inspect later.
Fan out one question, critique a document, review a plan, or force a strict option vote across validators. Fast cross-model truth without a full sprint.
A/B compare proposals, policies, PRDs, launch copy, or files by criteria. Pipe logs or notes into every model when you want independent reads.
Add, remove, list, and smoke-test validators from the CLI. Built-in adapters plus custom subprocess and HTTP providers. No hand-editing JSON.
Run objective evidence gates: tests, lint, coverage, builds, custom score scripts. Save baselines and block regressions before validators even argue.
Persist recurring findings, search hotspots, and run council-style deliberations for ADRs, PRDs, RFCs, launches, budgets, migrations, build-vs-buy, and incidents.
$ brew install traylinx/tap/makakoo $ makakoo install
$ curl -fsSL https://makakoo.com/install.sh | sh $ makakoo install
> iwr https://makakoo.com/install.ps1 | iex > makakoo install
Public beta: macOS arm64/x64, Linux arm64/x64, and Windows x64. Windows ARM64 is not published yet. Binaries are unsigned/not notarized while signing is being set up, so expect OS security prompts. MIT licensed. No telemetry. No cloud requirement. No vendor lock-in. Your machine stays yours. A few minutes from zero to Brain + AI bodies connected.
Makakoo is not a startup. Not a company. Not a SaaS. Not a closed platform. It's an open project built by contributors from everywhere. MIT forever. No VC. No acquisition path. No telemetry. Two funding rails at launch: GitHub Sponsors + Open Collective. Every contribution category is first-class - code, docs, mascots, design, translation, ideas. Every mascot has a named maintainer.