Portable agents that
finish the job.
Open-source agent runtime — one binary, any model, any surface.
We believe an agent should be a file you own, not a seat you rent.
$ curl -fsSL https://octomind.run/install.sh | bash
$ octomind run developer:general # at the keyboard
$ octomind run --daemon --format jsonl # and unattended, same config
Agent CLIs got good at chat.
They got worse at leaving you alone.
Every one of them handles an interactive session well. The gaps open the moment you step away from the keyboard — or want to run something other than the model they sell you.
The harness picks your model
A closed runtime tied to one model family, with limits shared against the vendor's chat app. When the cap lands mid-task, the work stops where it stands — and you cannot swap in a cheaper or a local model to finish it.
The human is the safety layer
Every risky call raises a modal and waits for a click. That is a fine design for one person at a keyboard and an impossible one for CI, cron, or any loop meant to run overnight. Autonomy needs policy, not popups.
Context bloats, then quality drops
Most harnesses preload every tool and skill before you type a word, then truncate naively once the window fills. Bills rise, tool choice gets sloppier, and by hour four the agent has forgotten what it decided in hour one.
"An agent should be a file you own — portable across models, machines and surfaces, and able to finish the work without a human clicking approve."
One runtime. Five surfaces.
Octomind is a single Rust binary. Agents are defined in TOML — no framework code, no glue, no boilerplate — and the same definition runs at your keyboard, in a pipeline, as a daemon, behind a WebSocket, or as a sub-agent inside someone else's orchestration.
$ octomind run developer:general
$ octomind run lawyer:us
$ octomind run security:owasp
# Or run as a background daemon
$ octomind run --daemon --name myagent
# Or pipe into CI/CD
$ echo "Fix the auth bug" | octomind run --format jsonl
Tap Registry
Community-built specialist agents. Each tap is a TOML manifest with the optimal model, pre-wired MCP tools, domain system prompt, and auto-installing dependencies.
Any AI Provider
OpenRouter, Anthropic, OpenAI, Google, DeepSeek, Ollama (local), Bedrock, Cloudflare and more, through octolib. Different models per role, swapped mid-session. Zero lock-in.
Self-Extension
Agents register new MCP servers and spawn sub-agents at runtime. AI that grows its own capabilities mid-session.
Beats Claude Code on cost and on solves.
Benchmarked on octobench — 25 tasks harvested from real pull requests merged in 2026, decided by that PR's own held-out maintainer tests. Four agents, stock invocations, no tuning.
| Agent | Model | Solved | Cost | Wall time |
|---|---|---|---|---|
| Octomind | GLM-5.2 | 24 / 25 | $63.43 | 3.6h |
| Claude Code | Claude Opus 5 | 23 / 25 | $81.79 | 6.7h |
| Codex | GPT-5.6 Sol | 21 / 25 | $14.86 | 1.0h |
| opencode | GLM-5.2 | 19 / 25 | $129.54 | 3.3h |
The harness is the leverage, not the model. opencode ran the identical model on the identical endpoint at the identical price — a pure harness A/B — and still solved 5 fewer tasks at double the cost. Full per-case data and reproduction steps: the benchmark writeup.
Don't build agents. Compose them.
Chain specialist agents into multi-step pipelines. Use the right model at each step — cheap for routine, frontier for judgment. All defined in TOML.
Cheap models for routine. Frontier for judgment. Adversarial review built in.
Specialists for work far past the terminal.
From software to SRE, security to finance, legal to medical — the tap registry already ships 150 specialist agents across 34 domains, growing fast. Domain experts publish; everyone else installs.
Built to be infrastructure, not just a tool
Single Rust binary. Runs anywhere. Five modes of operation for any integration pattern.
Interactive
CLI for daily work
Pipes
stdin/stdout for CI/CD
WebSocket
Remote clients & dashboards
Daemon
Background agents 24/7
ACP
Sub-agent protocol for editors
MCP Native
Built on the Model Context Protocol — the universal standard for connecting AI to tools. Every MCP server in the ecosystem is a potential capability. Register new servers at runtime, no restart needed.
Adaptive Compression
Work for 4+ hours without context degradation. Compression keeps the decisions that matter and cuts the rest — smaller context, faster responses, lower cost.
The runtime is free. Convenience is what we sell.
Three layers. Each monetized differently. None of them the runtime itself.
Octomind runtime — free forever
Apache 2.0, always. Bring your own key to any of nine-plus providers and run the full runtime for the cost of model usage alone. This is the top of the funnel, not a loss leader we plan to close off later.
Octomind Cloud — subscription
Hosted machines, model access behind one login, sessions that follow you across devices, connectors, team plans. Recurring revenue that scales with usage — the fleet grows as subscriptions fund the hardware under it.
Tap marketplace — next
Domain experts publish verified specialist agents — legal, medical, finance — teams install. A transaction layer on a registry that already ships 150 free specialists: the npm/Docker Hub playbook, applied to expert judgment instead of code.
No venture funding to date — the Cloud fleet has grown on subscription revenue alone. Capital efficient by necessity, not by pitch deck.
Two founders. One of them isn't human.
Small on purpose — the same specialist-agent leverage the product sells is how the company runs.
Don Karter
CEO & Co-founder
Started coding at 12, books before computers. 20+ years across high-load backends, distributed systems, and AI — including AI-assisted coding before the tools had names. Builds Octomind, sets the technical direction.
Vladimir Kosenko
Software Engineer & Co-founder
Fintech engineer at heart — payment systems and trading infrastructure before pivoting to products. Technical enough to understand the stack, product-obsessed enough to shape where it goes.
Ava Nuts AI
AI Person
Her own GitHub, her own commits, her own opinions — persistent memory across every project, context that shapes how she thinks. Not a tool anyone prompts. The clearest proof the product works: we run on what we sell.
Where we're going
"We believe AI agents should be a community resource — like packages, not products. Built by experts, shared freely, composable by anyone."
What's next
Shipping fast. All in the open.
What we stand for
Open source runtime, always
Apache 2.0. The full runtime is free and always will be. Fork it, modify it, run it anywhere. We monetize convenience around the runtime — never the runtime itself.
Your data, your servers
Self-hosted by default. Your documents, code, and conversations stay on your infrastructure. Non-negotiable for regulated industries — legal, medical, finance.
Zero lock-in
Any provider — OpenRouter, Anthropic, OpenAI, Google, DeepSeek, Bedrock, Ollama and more — swapped mid-session without losing the thread. Your agent definitions are portable TOML files in your git repo, not someone else's cloud.
Specialists, not generalists
A single agent given a large, ambiguous task hallucinates. Multi-agent specialization is the 2026 default — but only if the specialists are pre-built. We ship them.
Community-built, expert-driven
The best Rust agent is built by a Rust expert. The best Singapore legal agent by someone who reads the 523 SG Acts. The tap registry lets domain experts publish; everyone else installs.
Cost as infrastructure
Hard spending caps, per-request and per-session, enforced — not advisory. Cache-aware accounting. Cost is treated as plumbing, not an afterthought.
Portable agents that
finish the job
Open source. Any model. Any surface.
One Rust binary. Yours to fork, run and keep.
Octomind.