Quickstart
Build an agent, publish v1, run it, change it, diff the two. About four minutes.
This walks the whole loop the product exists for: a version you can point at, a run you can reproduce, and a diff that tells you what your edit actually did.
1. Create your workspace
Sign up at /signup with a name, email and password. You get an owner account, a workspace of your own, and one starter agent published as v1 so there is something to pin a thread to straight away.
Everything — agents, versions, threads, runs, files, keys — belongs to that workspace. Teammates are added from Settings; roles gate publishing a version and moving the stable pointer, and nothing else.
2. Check your credit
Runs use Coldcache's own provider accounts with Claude, ChatGPT and Gemini by default, metered against a credit balance you top up from Settings — there is no key to configure. Enterprise workspaces can run on their own provider keys instead; if that applies to you, Settings shows a key form in place of the balance.
3. Create an agent
New agent asks for a name and gives you five panes: instructions, tools, output, model and memory. The defaults are deliberate — web search on, memory off, temperature 0 where the model accepts one.
Write the system prompt as if the model has never met you, because it hasn't. Every run starts from an empty context.
4. Publish v1
Publishing freezes the body and assigns the number. From this moment the configuration is retrievable forever, and any chat that pins to v1 will behave the same way in a year.
5. Start a pinned chat and ask something
The chat header shows the pin — agent slug plus version number — for the life of the chat. It never changes. The answer streams with citations, and the run record below it carries the input hash, output hash, latency, tokens and the model id that actually served it.
6. Attach a document, or freeze one into the version
Attach a file to a question with the attach control in the ask box — it is hashed into that run's input hash. To make a document part of the agent itself, add it under Reference documents in the editor: it is then read on every run and its hash becomes part of the version's identity.
7. Edit, publish v2, diff
Edit opens a draft. Change the prompt, publish v2 with a changelog line, then use Rerun on another version from the original run. Coldcache forks a new chat pinned to v2, replays your input verbatim, and drops you on the side-by-side diff.
What you should have now
- Two published versions, both permanently runnable.
- Two runs on the same input hash, one per version.
- A diff that attributes the change in output to the change in configuration.