AI transformation infrastructure · Self-hosted · Any model · Apache-2.0
Your knowledge base, agents and workflows live on your own server, not inside an AI lab's product. Run Claude today, GPT or Grok tomorrow, or an open-weight model on your own hardware when the price is right. Run it in-house, license it as a consultant, or have us set up your server and run it for you.
Bring your own agent. Keep your own work.
Agent-agnostic
OperationKit doesn't ship a personal agent of its own. Point whichever one you already run at its MCP server (read-only: search, read, and list your knowledge base and board, no write access) and it drives the work.
Official MCP client. Bundled into Claude Pro/Max; add OperationKit with claude mcp add.
Official MCP client, stdio and Streamable HTTP. Bundled into ChatGPT Plus/Pro/Business.
verified: MCP clientMCP servers work out of the box, including headless/scripted use, on every X Subscription tier.
verified: MCP clientNative MCP support, tools/resources/prompts. Free, MIT-licensed, self-hosted, bring your own LLM key.
verified: MCP clientSupports MCP only through its own prebuilt connectors (Notion, Atlassian, Linear). No documented way to add a custom server today, so it can't drive OperationKit directly yet.
Not yet compatibleRun your own coding subscription, not a shared one: Claude, OpenAI, and xAI all permit automating your own single account and all prohibit pooling or reselling one seat across different people. If you run OperationKit's workspaces for genuinely separate clients or teams, give each its own account, not a shared seat.
Why own it
Here's how it usually goes. You build on one lab's models, agents and memory. You get productive, then you depend on it. Then the price goes up, and your playbooks, context and history are stuck inside someone else's product.
Operation Kit keeps all of that on your server, as plain files and a database you control. The model underneath is swappable. When one gets expensive, gets worse or goes down, you switch and keep working.
Own your data. Skills, agents and knowledge are files you keep.
Own your compute. Your server, with no limits set by someone else.
Use any model. Claude, GPT, Grok or open-weight, and you can change whenever you want.
Any model
OperationKit is the car: it holds your workspaces, your knowledge base, your agents, your history. The model is the gas station you fill up at. You don't rebuild the car to use a different station; you just pull into whichever one has the best price today, or is actually open.
Native engine. The default route, with an Opus fallback if a pinned model goes unavailable.
native engineNative engine. GPT-style model ids route automatically to the Codex engine.
native engineNative engine. Grok ids route automatically to the Grok engine, as a real job on a VPS you control.
native engineRoutes through the bundled LiteLLM proxy. Powers the research model group today, with a Claude fallback.
via litellmRoutes through LiteLLM to a model on your own GPUs. Powers the code model group today, with a Claude fallback. The concrete proof that "your own hardware" isn't aspirational.
via litellmEach model group is independently configurable. Today's default config routes most reasoning and writing work to Claude; wiring the rest is on the roadmap.
RoadmapvLLM and similar inference servers aren't wired in yet. Not promised until it ships.
RoadmapWhat transfers when you swap: skills, agents, the knowledge base, and workspace state. All of it is plain markdown or a database you control, never stored inside a model provider's product.
The field left something open
Hermes, OpenClaw, and Grok Bot each install in one command. Paste your OperationKit API key, or point an MCP client at it, and they drive objectives and read your team's shared knowledge base. Running more than one coding subscription of your own? The account router rotates across them for you, your accounts, your rate limits, never a seat shared across different people.
Six-phase pipeline, C-suite personas, per-workspace context. Plain markdown you own.
One deployment, isolated workspaces, real RBAC. No sales call required.
An adversarial reviewer grades every output against locked criteria.
Under the hood
Tools → Skills → Agents. Edges declared in frontmatter, validated by okit. Dangling references are build failures.
Every objective runs in its own dedicated session. Strategy objectives spawn workers and re-wake on completion. Workers run, get reviewed, report back. SQLite-backed state survives restarts.
A fresh-context reviewer grades output against pre-locked criteria, blind to the worker's scratch. Auto-iterates up to three times. You see results, not drafts.
Four tiers load before the model sees a token: global routing, agent persona, workspace context, user profile. Late-loaded tiers cannot override earlier rules. OperationKit has four, validated by okit validate.
53 tools. 82 skills. 19 agents. Every edge declared in frontmatter. okit validate hard-fails on any slug that does not resolve. A renamed tool is a build error.
Honest answers
Hermes, OpenClaw, and Grok Bot already install in one command.
They do, and OperationKit doesn't compete with them, it's what they drive. Give any MCP-capable agent your OperationKit API key and it controls objectives on the board and reads your knowledge base live. OperationKit adds the workspaces, shared context, and a router that spreads your own accounts' rate limits, never someone else's seat.
What about my API keys?
It runs on your server, code you can audit. Your keys do not leave your machine. Not a hardened managed SaaS, not a security guarantee. If "I need to read every line that touches my keys" is the bar, this is it.
How do I evaluate this before committing?
Clone github.com/m1keluka/OperationKit. Start with SECURITY.md; it states the threat model. Apache-2.0. Star the repo to follow releases.
The source is live. Star the repo, or leave your email for release notes.
Or star the repo to follow along.