The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Indraft listing page.
A CRM your AI agent can actually run. Connect Claude, ChatGPT, Cursor or any MCP client and it can read and write companies, contacts, opportunities, tasks and interactions from the first request, with no schema to design first.
MCP server: https://api.indraft.io/mcp · indraft.io ·
Docs
Claude opens a browser to authorize. Nothing to paste, no API key to store.
In Claude Desktop or claude.ai, add it as a custom connector with the same URL.
Settings → Connectors → Add custom connector, URL https://api.indraft.io/mcp.
It authorizes over OAuth.
Any client that speaks Streamable HTTP and OAuth 2.1 works. Point it at
https://api.indraft.io/mcp.
Then paste this and your agent will set itself up:
Set up Indraft for my business. Read its CRM schema first, then tell me what you think we do and how you plan to model it, and wait for my answer before you write anything.
It reads the schema before it writes, so it learns the object types, the limits and what it is allowed to do without you explaining any of it.
Most CRMs were built for a person clicking through forms, and an agent using one is doing screen-scraping with extra steps: guessing field names, inventing records, and overwriting things somebody already checked.
Indraft is built the other way round.
Every value records who asserted it and how. Observed, inferred, imported, supplied by a person, or confirmed by a person. Those rank, and a weaker claim cannot quietly overwrite a stronger one, so your agent's guess never overwrites what you checked. A write that was held says so rather than reporting success.
Every change is on the record. The actor, the surface it arrived through, and the field-level diff. Most changes can be undone.
It is populated from the first request. Companies, contacts, opportunities, tasks and interactions exist before you configure anything, with a pipeline that can actually close a deal. Custom fields and your own object types are there when you need them, defined by asking.
It refuses rather than guessing. A total that crosses currencies is refused unless you have stated a rate, and every converted answer names the rate it used and who asserted it. There is no bulk delete.
45, covering reading, writing, schema changes, merges, bulk imports and aggregation. An agent is only shown the ones its credential can actually call.
Full reference, generated from the server's own catalogue: TOOLS.md
Every tool carries a human-readable title and the applicable readOnlyHint,
destructiveHint and openWorldHint, so a client can tell a read from a write
before it calls one.
For your own service, when you want the REST API rather than MCP. The source is
in sdk/; an npm release follows.
Errors are typed and carry a recovery hint rather than a status code alone:
A cross-currency total is not an error, it is a different answer. Ask for a sum over deals in several currencies with no rate on file and the aggregate comes back grouped per currency rather than summed, because a total that silently mixes currencies is the failure this product refuses to commit.
See sdk/ for the source. MIT.
io.indraft/mcpIndraft is live at indraft.io. Free tier, no card.
Issues and questions are welcome here. Security reports go to
koshea@justicebid.com rather than a public issue.
MIT for the SDK and everything in this repository. The hosted service has its own terms and privacy policy.