How Orkka works

Most agent tools give you a chat box and a diff. Orkka gives you a board and a queue of decisions. The agents generate work continuously; your job is to steer and approve it. Every surface in the product exists to make that job faster.

The practical consequence: you can run several tasks at once without losing track of them, and you are never asked to trust a change you have not seen evidence for.

Tasks move left to right through lanes. Agents own some lanes, you own others — and a task can always be moved backwards by hand.

Testing and Code Review are configurable per project, and any individual task can skip planning, testing or review when it genuinely does not need them.

Running five agents in parallel is easy. Reviewing what five agents produced is the actual bottleneck, so Orkka treats it as the main event.

Every project is seeded with the full roster. You can rename agents, change their model, and give them custom instructions; you cannot add or delete them. Which ones you actually use is up to you.

Agents have persistent memory per project and a full execution history — token spend, files touched, and the log of every run they have done.

Unmanaged projects use your own repositories, any stack. Orkka detects each repo's stack and assigns appropriately specialized dev agents — no frontend agent on a Go service.

Managed projects are for non-technical users: Orkka creates the repository — Next.js + Tailwind, with Prisma and a Postgres database when the app needs one — and deploys it live at its own orkka.app address.

Agents work in a bare clone plus one ephemeral worktree per task, so parallel tasks on the same repository never collide. Orkka detects dependencies between tasks — automatically as well as manually — and holds a task until whatever it depends on has merged.

Orkka drives the vendor CLIs rather than reimplementing them: Claude, Codex (ChatGPT), Gemini, Grok (xAI), GitHub Copilot, the open-source OpenCode harness, and local models via Ollama. The QA browser-testing lane currently requires Claude.

The last two are worth distinguishing, because they are opposites. Ollama serves models but cannot edit files or run commands, so Orkka drives it through an agent CLI. OpenCode is the reverse — a coding agent that ships no models of its own and runs against whichever providers you signed in to with opencode auth login. Its models therefore aren't a fixed list: your runner reports what your logins actually reach, and those become your choices.

Runs execute in one of two places. On your own runner, everything happens on your machine using the AI subscriptions you already pay for — no per-token cost, and your code never leaves your computer. In the hosted cloud, Orkka runs the agents with its own keys and bills usage plus a 10% fee.

Orkka Local is available today and free. The hosted cloud is live too: sign up and your team runs on Orkka's computers, with 10,000 free credits to start.

Agents write code; the platform handles infrastructure. They never receive database credentials, deployment tokens, or payment keys. Bash access is guarded, dangerous commands are blocked, and every command is audited.

Each run is bounded by turn, tool-call and wall-clock budgets, so a confused agent stops rather than looping — and a task that hits a limit escalates to you instead of failing silently.

Every project has a memory the whole team shares: decisions and their reasons, conventions of the codebase, preferences you have expressed, and what agents learned (or got wrong) on earlier tasks. Agents save to it with a remember tool and search it with recall; each run also leaves a short outcome note. Nothing is injected wholesale — for every run and chat turn Orkka retrieves what is relevant: pinned entries and standing rules first, then full-text matches on the task, then recent learnings, under a fixed budget.

The Memory page lets you read, search, add, edit, pin, archive and delete entries, and “What would an agent see?” previews the retrieval for a task so you can pin a decision that isn't surfacing.

A skill is a SKILL.md your project keeps — a repeatable procedure such as grooming the backlog, writing release notes or running a security pass. Assign skills to agents on the Team page; assigned skills are written into the agent's worktree and listed in its prompt, so it reads the full procedure only when the task matches. In chat, type /skill-name to run one as a playbook.

Orkka is an MCP server. Mint a personal API key under Settings → API keys (it acts as you, with the permissions you pick, optionally pinned to one project) and point any MCP client at /mcp with the key as a bearer token:

claude mcp add --transport http orkka https://api.orkka.ai/mcp --header "Authorization: Bearer <YOUR_KEY>"

Tools cover projects, tasks, plans (approve / reject), reviews (approve / merge), questions, autopilot, skills, schedules and agent chat. The same key works against the REST API. Every write through MCP lands in the audit log.

Every management action (members, permissions, settings, agents, skills, schedules, budgets, API keys, webhooks, Slack) and every agent run is recorded with who did it. Entries form a per-organization SHA-256 hash chain; Settings → Audit log lets you filter, verify the chain and export CSV/JSON — the export is itself audited.

Install Orkka Local and create your first project — the installation guide covers prerequisites, the install one-liner, and troubleshooting. Start with one small, well-specified task on a repository you know well: it is the fastest way to calibrate how much you need to say for an agent to get it right.