DIN is the AI work-of-record. It captures what an AI agent actually did at the tool-call boundary, grades each claim as evidence-supported or asserted, and keeps that record on infrastructure you control. It is for teams whose agents do real work. DIN is prelaunch: the free build is one field and one AI, with no signup and no card.
A log tells you the agent ran. It does not tell you whether its claims survived contact with the evidence. A transcript of a claim is still a claim, written in the same confident voice whether the work held or not. DIN keeps the grade: what was proved, what was only stated, and what was later contradicted.
The record lives in a field: one canonical store per tenant, where records are placed by what they rest on rather than filed by date. It runs on a machine you control, and the field listens on that machine only: agents on the same machine connect through the plugin or the CLI, cloud chat history comes in through import, and hosted access is on the waitlist.
Five steps, the same for every producer, whether the work came from a Claude Code session or a Codex rollout:
Two questions, asked separately and never merged into one badge: how a record was obtained, and what it turned out to be worth. How it was captured is one of receipted, logged, inferred or mixed. What it proved is one of proved, asserted or refuted. A refuted claim stays in the record, marked, rather than being quietly removed.
An inferred record never renders as a receipted one, anywhere in the console, at any zoom. Where the record cannot settle a position on its own, a person rules, with a name and a date on the stamp.
One meter: fields. A field is a store, a team, a project, a client: one container of records, opened by one token. You are not billed for volume, for seats, for storage or for inference, and your model calls run on your own key, so there is no usage line.
The free build is one field and one AI, with no signup and no card, and it is available to download today for Linux x86-64 and Windows x86-64.
Paid plans are priced by number of fields, in USD per month. Self-hosted: Starter $29 (6 fields), Freelancer $99 (15), Team $199 (30), Small Business $299 (90), Agency $399 (90 fields across up to 5 orgs), and Enterprise I to V from $699 (180 fields) to $3,499 (1,000 fields). Hosted, where we run it: from $50 (Starter) to $4,999 (Enterprise V). Above 1,000 fields is quoted. DIN is prelaunch and no paid plan is on sale yet: the pricing page takes a waitlist signup, not a card.
DIN grades claims against the evidence that produced them, so the evidence must still exist. Eight sources import and grade today: Claude Code, Codex, Gemini CLI, Qwen Code, the Cline family, claude-mem, Grok, and Zep / Graphiti (through din-import-graphiti). Other tools that keep the original conversation (Letta, LangGraph, OpenAI Sessions, AWS AgentCore events, Azure Foundry conversations, Google Agent Engine, Supermemory, Khoj, Basic Memory, SpecStory) can be graded once a mapper for them ships. ChatGPT import is not supported yet. Tools that store only extracted facts cannot be graded, because the thing a grade would point at is gone.
Live capture: the installer asks which agent to capture, Claude Code, Cursor or OpenCode (or takes --agent). Codex, Gemini CLI, Qwen, Cline and Grok are import-only.
The free build is: one field, up to 3 people on one AI account (one AI at a time), with graded history import, live capture and the local dashboard, no signup and no card, running entirely on your own machine. Paid is the same download plus a key that turns on several agents on one field, more fields and the paid panels. DIN is prelaunch.
On infrastructure you control: your laptop or your own box. The installer starts one server on loopback behind a token it mints on your machine, and the field listens on that machine only: agents on it connect through the plugin or the CLI, and cloud chat history comes in through import. Model calls run on your own key or your own local claude.
Linux x86-64 and Windows x86-64. There is no macOS binary yet, and the installer says so rather than failing halfway through a copy.
The Community Edition, the plumbing, is licensed Apache-2.0; its public repository is not published yet. The compiled grading engine and the commercial services are proprietary.
No. The console is real and running, and every count, trust mark and flag is read live from a field. The records in that field are invented: Northwind Trading Co. is a fictional company, nothing here is anyone's real data, and the console is read-only.
Download the free build · Pricing · Docs · For AI agents
Updated 2026-09-30.
A name and an address. No password, no card, and nothing to install before you have decided.
Real: this console, the field behind it, and every number on every screen. Nothing on this page is a mock-up, a screenshot or a canned response. Each panel is a live read of a running FieldDIN field, and a panel whose endpoint does not exist says so instead of drawing something plausible.
Not real: the records in that field. Northwind Trading Co. is a fictional company. Calloway Foods, Tidewater Mutual and Brightline Rail are fictional clients. The four people who appear as authors and reviewers are invented, and so is every path, command, test output and refutation. The data set is generated from a scenario file and checked, on every build, for anything resembling a real name, host or record.
Read-only. This console proxies GET requests to a fixed list of read routes and holds no path to the field's one write door. Nothing you click can change a record, and nothing you type leaves the page.
This is FieldDIN, running, reading a live field over the network and grading what it finds. The grading is the product, so none of it has been softened for a demo. What has been made up is the work being graded.