Use Dhanur AI from Claude Code (MCP)

Dhanur AI is an MCP server. Connect Claude Code, Codex, Cursor or another MCP client, and you can ask it to create an agent, give it knowledge, publish it, talk to it and read what it did, without opening the app.

Connect

Create a project API key: Integrations, then API keys, then Create a key. Owners and Admins can do this. The key reaches only that project.
Add the server to your client. For Claude Code:

claude mcp add --transport http dhanur https://engine.dhanurai.com/mcp --header "Authorization: Bearer dhk_…"

Other clients take the same two things: the address https://engine.dhanurai.com/mcp (Streamable HTTP) and the header Authorization: Bearer <your key>.

The dialog that shows a new key also shows the Claude Code line, ready to copy.

What it can do

Tool	What it does
`list_agents`, `get_agent`	The project's agents, and one agent's draft settings.
`create_agent`	A new draft agent from a name and instructions, with knowledge or a template if you like.
`update_agent`	Change a draft's name, description, instructions or knowledge.
`add_knowledge`	Add a knowledge note to the project, and to an agent's draft.
`publish_agent`	Make a draft live, with the same checks as the app.
`run_agent`	Send a message to an agent's live version and wait up to 30 seconds for the reply.
`get_task`, `list_tasks`	Tasks started through the API, schedules, webhooks and events.
`list_approvals`, `decide_approval`	Actions waiting for a person; approve (a reply can be edited first) or reject.
`list_test_sets`, `run_test_set`, `get_test_run`	Run an agent's test sets and read the results.
`usage`	What the project's agents cost over the last days, in rupees.

Try asking your client: "Make a Dhanur AI agent that answers questions about my bakery's timings and delivery, add these facts, test it with three questions, and publish it if the answers are right."

Rules

One project. A key reaches only its own project.
Changes are made as the person who made the key, with their role in the workspace now. If they leave, or are no longer a Builder, the key can still read and run agents, but not change them. Every change is in Settings, then Activity, marked as made through the API key.
The same checks as the app. Publishing refuses our example shop's facts and waits for failing test sets; there is no override through a key. Runs use the published version, and your approval settings, spending limits and credits apply.
Limits: the same per-key limits as the run API: 30 runs and 240 other calls a minute.
Costs: runs and test runs are charged like any others. Using the MCP server itself is free.

Revoking the key under Integrations cuts the client off at once.