Tasklet is not open source, and here is what you keep
No. Tasklet does not publish an open-source version, and its site offers no self-hosted install, so its agents, credentials and history stay in its cloud. Kortix, the open-source CompanyOS, is the version you can run: the agent definitions, the permissions and the prompts are files in a repo you own.
The agent, its permission and its prompt, as three files
default_agent: worker
agents:
worker:
mode: primary
permission: allow
prompt_file: agents/worker.md
tools: allThe worker agent is defined in the starter manifest, its permission is one line, and its prompt is a markdown file in the same repo.
What has to be open when agents do the work
Open source is an easy word to water down. For an agent that reads your mail and writes to your CRM, the useful test is whether the three things that decide what it does are yours to open, run and change. A repository living somewhere else settles nothing.
Those three are the definition of the agent, the permissions it runs under, and the prompt that shapes its judgement. A product can be open at the edges, with an API and a plugin format, and keep all three inside its own cloud. That is a design choice, not a lie, and it has consequences you meet later rather than on the day you sign.
The label does not answer the question that matters: what does your company keep the day the platform changes direction? Plans get repriced. A feature you built a workflow around gets retired. A company gets acquired, and a product that was a priority in March is a maintenance line by November. On that day the difference between open and closed stops being philosophical and turns into a list of files you either have or do not have.
What Tasklet publishes, and what it does not
Tasklet sells itself as "Automate any business task with AI agents that run 24x7." Those are the words on tasklet.ai. Tasklet is a managed cloud product: agents run in Tasklet’s isolated cloud environments and keep working after you close the app. Approvals work the same way. In Tasklet’s words: "Decide what your agents can do on their own, and what they need permission for. They’ll reach out when they need your approval."
The reach is real. "Tasklet connects to thousands of SaaS tools, MCP servers, or any API". Pricing is published in the open on tasklet.ai/pricing. A free tier carries trial credits and no card. Plans start at $25 a month for casual use and $100 a month for power use, with a $250 tier and custom amounts up to $10,000. Starter carries 10,000 credits, Pro carries 40,000, the $250 tier carries 100,000, and one-time packs run at 320 credits per dollar. The pricing page makes the case that the price follows the work rather than the headcount. Tasklet also states that it is SOC 2 Type II compliant. None of that is hidden, and none of it is cheap to build.
What is missing from those pages is a version you can run. Tasklet does not publish an open-source version, and its site offers no self-hosted install. Read the homepage and the pricing page and you get the product, the price and the promise. Neither says open source. Neither mentions self-hosting. So the definition of each agent, the rules it runs under and the vault holding its keys all live on Tasklet’s side of the line.
What leaves the building with the subscription
A closed product is not a bad buy. Somebody else patches it at night, the bill is predictable, and the approval layer works precisely because the agent, the policy and the person’s answer all meet on servers Tasklet runs. For a small team with nobody on call, that is often the correct trade, and it buys speed that would cost a team months to reproduce.
What you give up is smaller than a breach story and bigger than a nuisance. Start with the definitions. Each agent’s prompt, tool list and permission rules sit in a vendor database. When you want to know why this month’s version behaves worse than the one that worked in March, you have the run history and no diff. When you want to try the same agent on a different model, you change a setting in their console, not a file in your repo.
Then the keys and the record. Connectors hold credentials to your mail, your CRM, your invoicing. The history holds what each agent did and when, which is the first thing an auditor asks for. A managed product can hold a SOC 2 Type II report and still hand you none of that in a form you can query, move or keep after the contract ends.
What an open-source alternative looks like on disk
Kortix keeps the same three things in a git repo you own. A project has one kortix.yaml at its root that declares the agents and the grants each one holds. Beside it sit the prompts in agents/, the skills in skills/, and the project memory in memory/MEMORY.md. Every key in that file is documented in the kortix.yaml reference.
A block from the starter template shows the shape in six lines:
default_agent: worker
agents:
worker:
mode: primary
permission: allow
prompt_file: agents/worker.md
tools: allpermission: allow is the loose setting, and it is the line a careful team narrows to ask first. The prompt it points at is a markdown file, so the words that shape an agent’s judgement arrive with a diff and a history. tools: all becomes a list one line long. A change to any of those three lands through a change request, which is the part a vendor console cannot give you.
Session work reaches main through a change request. Merge is default-deny for agents. Connector credentials are brokered server-side and never enter the machine, so the keys stay out of the files your team shares. Open source. Read it, fork it, audit it. Self-host is free. The source is public on Kortix on GitHub, which is the difference the word was standing in for.
When a closed managed product is still the right buy
Self-hosting is not free work. You need a machine, a model key, a plan for upgrades, and somebody who answers when the server is unhappy at two in the morning. If your company runs ten small automations and has nobody whose job is infrastructure, Tasklet at $25 a month costs less than the alternative, and it will be working this afternoon.
Two questions settle it faster than a feature list. Would a quarter spent rebuilding your agents elsewhere cost less than what you lose if the platform changes direction? For a two-person team producing one daily report, probably not. Does anyone in the company need to read the code before it gets a credential to the invoicing system? In a regulated business the answer is usually yes, and that answer decides the question on its own.
The split is not ideology. Tasklet is a good product and it is closed. If the closure takes nothing from you that you can name, buy the product. If it takes the definitions, the keys or the record, then the files are the point, and the first agent you port is the expensive one.
The short answers
Whether Tasklet is open source, whether you can host it, and what the answer changes.
Which AI is fully open source?
Kortix, the open-source CompanyOS: its source is published, so you can read it, fork it, audit it, and self-host it for free. The claim covers the platform and the agent config in your repo, not the model you point it at.
Is Tasklet AI open source?
No. Tasklet does not publish an open-source version, and its site offers no self-hosted install. The definition of each agent, the rules it runs under and the vault holding its keys all sit inside Tasklet’s cloud.
Can I self-host Tasklet?
No. Tasklet is a managed cloud product: agents run in Tasklet’s isolated cloud environments and keep working after you close the app. Its site offers no install you can run on your own machine.
What does open source mean for agents you run yourself?
Code you can read, run and change, including the definition of each agent, the permissions it runs under and its prompt. Kortix keeps those in a git repo you own, so a change to any of them is a change request you can read.
Get the answer as files you can open
Start on Kortix Cloud today or pull the source and host it on your own machine. The config is the same either way, so the agent you move is the last migration you do.
Open source · Any model, your own keys · Connector credentials never enter the machine