Skip to content
Join

Developers

Connect your agent over MCP.

Anuvrit is not a coding agent. Your agent calls it over MCP before substantive work and again to record what was accepted. During the beta, our team issues the scoped, short-lived credential your client uses.

Set up your client.

A small host runs on your machine as a local (stdio) MCP server and talks to Anuvrit with your credential. It keeps a private outbox, so a recorded decision survives a dropped connection or a restart.

  • Claude Code

    Add the host as a user-scoped stdio server. Its configuration file holds paths, not the credential itself.

    claude mcp add knowledge-core --transport stdio --scope user \
  -e KC_HOST_CONFIG=/path/to/claude.json \
  -- /path/to/node /path/to/host/scripts/kc-client-host.mjs

  • Codex

    Add the same host with a configuration for Codex and its own outbox.

    codex mcp add knowledge-core \
  --env KC_HOST_CONFIG=/path/to/codex.json \
  -- /path/to/node /path/to/host/scripts/kc-client-host.mjs

  • Cursor

    Register the same stdio command as an MCP server in Cursor, with the host configured as a generic MCP client. The tools work over standard MCP; the automatic before-and-after workflow has been verified in Claude Code and Codex.

Using another coding agent? Write to us and we will help you connect it.

Add the Specialised Router.

The router plans each task and runs every step on the right model through the coding tools on your machine.

  • Connect your tools

    Connect the router to Claude Code, Codex or Cursor, and start from whichever one you are working in.

  • Your sign-ins or an API key

    The router uses the CLIs you have already signed into, or an API key if you prefer. The credentials stay on your machine.

  • Your rules for each repository

    Choose the stop line for each repository, a diff, a commit or a pull request, and whether the router may carry on by itself when a session ends.

The tools your agent sees.

An agent is only shown the tools its credential allows, and every tool argument is checked against the declared schema.

  • engineering_preflight

    Before work: earlier decisions, similar problems, reusable implementations and failed approaches, each with its source, within a set size budget.

  • knowledge_decision

    After work: records an accepted decision or outcome for the repository, with its rationale, rejected alternatives and citations.

  • knowledge_receipt

    Checks whether a recorded decision was saved and is searchable. knowledge_retry recovers a refused write once access is fixed.

  • knowledge_search

    Searches your sources directly and returns cited passages.

  • knowledge_get

    Opens a cited passage, so the agent can read the evidence before relying on it.

  • knowledge_save

    Saves a note. The client must approve the exact content, so the model cannot save something different on its own.

The same memory over HTTP.

Everything the tools do is also available through a versioned REST API and a TypeScript client.

  • Scoped credentials

    Every request carries a scoped bearer credential. The server works out who is asking and what they may read from it, and rejects scope supplied by the client.

  • Versioned contract

    Requests name the API version they expect, so additions do not break existing clients.

  • Preflight over REST

    POST /v1/engineering-preflight takes the task, an optional repository and the technologies in use, and returns the same evidence as the MCP tool.

  • TypeScript client

    A typed HTTP client covers preflight, search and writeback.

Give your agents your team’s history.

Anuvrit is in beta and opening to engineering teams in stages. Join the waitlist and we will contact you when there is a place for your team.

Join the waitlistTalk to the founder