Odysi CRM
Companies, people and deals, with the history of every value. Claude logs emails, writes notes, moves stages and creates tasks through 165 operations. Every change shows who made it: “Thomas ◆ via Claude” when his AI did.
A couple of cookies keep the site running; analytics help us see what’s useful. You choose what we keep, read the policy.
Until September 2026, Odysi Studio's pipeline was spread across Pipedrive, Notion, email threads and memory. Calls and client work sat in other tools again, so no single place said where a deal stood, what was agreed on a call or what a client was waiting for.
Claude could already read our email, Google Drive and calendar. What it lacked was a structured place to write what it found, under our rules and with each person's access.
What would a CRM look like if the Claude we already work in did the work in it, instead of an assistant built into the software?
Our first CRM, our website and some internal tools were built on Lovable. Setup was quick: the systems were connected for us and things mostly worked without much involvement. Building with Claude asks more of you.
But Lovable builds software with AI inside it. We wanted the opposite: the Claude we already work in, with its connectors to Gmail, Google Drive and Google Calendar and its browser, doing the work in our own software.
That meant writing an MCP server: a set of operations any connected Claude can discover and run without access to the code. Once it existed, anyone on the team could connect from their own Claude session and start reading, adding and changing records, for example updating a deal from a Gmail thread.
The first CRM we built on Lovable was a different product. We rethought it from the start in Claude Code, around the connector rather than around a screen.
We wrote separately about moving our website off Lovable: Claude Design instead of Lovable.
Odysi Suite holds and shows the data. The person's own Claude, or ChatGPT, does the work through the connectors.
Companies, people and deals, with the history of every value. Claude logs emails, writes notes, moves stages and creates tasks through 165 operations. Every change shows who made it: “Thomas ◆ via Claude” when his AI did.
The Odysi Recorder for Mac and Windows records a call on any call app. Scribe then works in three steps: AssemblyAI transcribes, Gemini names the speakers, and Claude Sonnet, through the Anthropic API, files the call on its meeting and deal and gives it a title.
A workspace per client project: tasks, messages, documents, approvals and results. Claude proposes and drafts through 161 operations; anything the client will read waits until a person publishes it.
The team's written rules, in a clear hierarchy: team documents, with each person's own rules on top. Every conversation and every scheduled run reads it first, from any session, and corrections are turned into new rules.
No built-in AI agent. Odysi CRM and Odysi Projects make no generative AI calls of their own, apart from suggestions in CSV import and duplicate checks. Odysi Scribe is the exception, for transcription. Everything else is triggered by a person, in their own Claude session or in a scheduled task.
Each person connects their own Claude once, with OAuth 2.1 and PKCE. Every call then runs with that person's permissions, each workspace has its own connector address, and the Playbook is read before anything else.
Each connector exposes five fixed tools. One lists the operations, two run them (read or write). New abilities reach open conversations and scheduled tasks without reconnecting, Claude's context stays small, and every write is explicit.
OAuth 2.1 with PKCE, one connector address per workspace, and every call made as the member who connected. Changes are labelled “via Claude” in Activity and on the record.
The Playbook is written once and read by every session: the whole team works from the same rules, in as many separate conversations as they like.
Claude stores emails, file links and calls in the record. The next task fetches exactly what it needs instead of reading whole Gmail threads again, so there is less context and the answers are more accurate.
Odysi Suite was built with Claude Code running Claude Opus, with Claude Design for the interface, Claude chat and Artifacts for specs, and Cowork scheduled tasks to run the studio on it while it was being built.
Work started on 19 September 2026 and Odysi Studio was using it every day within two weeks. By 18 November 2026: about 2,600 commits, about 320,000 lines of code, about 770 test files and about 200 database tables in PostgreSQL (Supabase), hosted on Railway.
Claude sometimes did not follow specific written rules, such as writing rules. Corrections given in a conversation were not always stored as rules unless someone asked, and it is easy to forget to ask.
We refined the instructions that turn corrections and feedback into Playbook rules until this worked reliably. The trade-offs remain: scheduled tasks need a paid Claude plan, and setup takes more effort than an all-in-one builder.
Claude does the reading, logging and drafting. The steps that reach other people, or that a client relies on, stay with a person.
Logging, updating and drafting happen in Claude, through the connector. We read, decide and send. Figures from Odysi Studio's own workspace, 18 November 2026.
Odysi CRM is in beta and other businesses can start on the Free plan. Odysi Projects is available on request.
A company can run its sales, calls and client work through Claude, with the software as the record.
MCP works best with stable tools, the person's own permissions and written rules that every session reads first.
Claude Code with Claude Opus took the suite from first commit to daily use in two weeks.
A small studio. We find where AI is worth it, build those few things and leave you able to run them.
Odysi Suite is how we run our own work. We build the same kind of system for clients: their Claude, working in software built around their data and rules.