Reply.io has shipped Jason Harness, a free, open source AI SDR that runs inside the AI coding agent you already pay for. The source opened on November 1, 2026, the team announced the launch to its audience on November 19, and three new releases landed in the last two days alone. There is no platform fee, no per seat price and no model API key. You paste one prompt into Claude Code or Codex, it installs a small runtime on your machine, and nine specialist AI roles plan, research, write, personalize and follow up on an outbound campaign. You keep the send button. Here is how to set it up and run a safe first campaign this week.
What is Jason Harness?
Jason Harness is a harness in the literal sense. It wraps the AI agent you already use and gives it what one chat session cannot hold: a small runtime on your machine that keeps campaigns, per person plans, approvals and an append-only history, so the work survives closed laptops, restarts and crashes. Nine specialist roles split the job the way a good sales team would: a planner, a researcher, a copywriter, a personalizer, a critic, a responder, an analyst, a deliverability lead and a campaign manager that watches the whole campaign. The project is MIT licensed with the full source on GitHub, built and maintained by the team at Reply.io.
The key design choice: Jason ships no model and holds no model keys. The thinking runs inside Claude Code on the plan you already pay for, within its usage limits, so a role can think twice about a draft without a per token bill growing. As of November 2026, Claude Code is the host Reply verifies against, Codex works through the same CLI and skills, and four other CLI agents are built in but not yet measured on real accounts. Jason is alpha software and changes fast, so expect rough edges.
What do you need before you install it?
- An AI coding agent on a paid plan. Claude Code on a Pro or Max plan is the verified host. If you already run parts of your outbound from Claude, you have the hard part done. We walked that path in our guide to running outbound from Claude.
- A sending provider. The Reply plugin ships ready: connect Reply once and email and LinkedIn work right away, with SMS and call steps inside sequences, a unified inbox and warm-up handled. Any other provider needs a plugin you or your agent writes.
- Node 20 or newer for the Reply CLI. Jason itself needs no Node, Python or Docker.
- A testing account. Reply's own docs say it plainly: approve real sends only in a provider account you keep for testing while the project is alpha.
How do you install Jason Harness?
The whole install is one prompt pasted into your agent:
Install Jason by following https://get.jasonharness.ai/install. When it reports Ready, start its quick-start skill and take me to my first campaign. If anything stops you, ask me directly and tell me why.
Your agent downloads one self-contained executable under your own account, checks the download's SHA-256, deploys the role skills and starts the runtime. No admin rights are needed. Reply measured the executable portion at 16 seconds on an Apple silicon Mac; the guided setup that follows takes longer. If you would rather not go through a terminal at all, Jason Desktop is a signed one download app for Windows and macOS that sets up the same installation. Either way, the command jason status confirms everything is ready, and each failed check names the command that fixes it.
How do you connect Reply and launch a campaign?
You sign in to Reply once with its CLI, so your API key never passes through the chat. Your agent wires up the plugin, its permissions and its routing from there. Then you brief Jason the way you would brief a new sales hire, in plain words: what you sell, who should buy it, which channel first. Contacts come from Reply's Live Data API, which spans more than 1 billion B2B contacts and previews the audience size before you spend credits, or from your own CSV, CRM or research tools. The first message waits for your approval, your feedback becomes a standing rule, and the sensible move is a pilot of ten people before you let anything scale.
What stops it from spamming people?
Until you say otherwise, every message that would reach a real person waits for your yes, with the full text, the recipient and the sending account in front of you. A critic role reviews drafts and sends back claims you cannot prove. A do-not-contact list never relaxes. A deliverability role reads bounces and opt-outs and pauses a campaign on its own when the numbers turn bad. You can pre-approve one narrow category of reply, with an expiry and a ceiling, and every automatic send is logged under the name of the person who allowed it. That is a stricter default than most hosted AI SDRs ship with.
What does it cost to run?
Jason itself is $0 and MIT licensed. The thinking is covered by the Claude or Codex plan you already have, and what you actually pay for is sending infrastructure through your provider. For comparison, Reply's own math puts a self assembled stack (data, email finder, mailboxes, warm-up, a sender and a LinkedIn tool) at around $280 a month at entry prices as of October 2026, across six logins, while hosted AI SDRs are mostly sold by demo on annual per seat contracts. If you are weighing those hosted options, see how the major AI prospecting agents compare, and note that OpenAI made its own sales move two days ago with the ChatGPT for Sales rollout. The AI SDR market is converging on the agent you already use, fast.
Frequently asked questions
Is Jason Harness really free?
Yes. The software is MIT licensed with no seats and no subscription, and it collects no usage data. You still pay for your Claude Code or Codex plan and for whatever your sending provider charges, so it is free software on top of infrastructure you were paying for anyway.
Do you need a Reply.io account to use Jason Harness?
In practice, yes, for now. Reply's plugin is the only provider integration that ships ready, covering email, LinkedIn, the inbox and warm-up in one account. Any other provider works through a plugin you or your agent writes against the published contracts.
Which AI agents can run Jason Harness?
Claude Code is the verified host and the one Reply measures real runs on. Codex works through the same CLI and skills. Cursor CLI, GitHub Copilot CLI, Gemini CLI and Antigravity CLI are built in from each host's documentation but not yet measured on real accounts as of November 2026.
Is it safe to let Jason Harness send emails on its own?
By default it sends nothing without your approval. Automatic sending exists but has to be switched on by a named person, passes a critic review, respects a daily ceiling and a permanent do-not-contact list, and is logged. The project is alpha, so run it against a testing account first.
Want an AI SDR wired into your CRM, your data and your deliverability by operators who build these every week? See The Prospector, our prospecting agent build service.