Skip to content

MCP server

AI coding agents (Claude Code, Codex, Cursor, Devin, anything that speaks MCP) can ask CodeIntent the same question the GitHub App answers on pull requests: what breaks in this repository if dependency X moves to version Y, and what exactly do I change?

The answers are deterministic. Both releases are compared statically against the repository's own code. Nothing is executed, and no model writes the brief.

Endpoint and auth

  • Transport: MCP streamable HTTP, endpoint /mcp. It is stateless, so any worker can serve any request.
  • Hosted endpoint: https://codeintent-api-ded80aad7f4b.herokuapp.com/mcp
  • Auth: every request needs Authorization: Bearer <token>. A missing or wrong token gets 401.

Create a token on AI agents. It belongs to one GitHub App installation, is shown once, is stored only as a SHA-256 digest, and can be revoked there.

Scope

With a workspace token, the installation's own repositories, private ones included, are read with its installation token. Any other repository is fetched as the public sees it. Each call downloads the repository tarball for the requested ref (300 MB cap) into a temporary folder and deletes it afterwards.

Tools

All four tools are read-only. Each returns structured JSON with an output schema. A failure comes back as a tool error with a short message the agent can act on, never a stack trace.

ToolArgumentsReturns
upgrade_briefrepo (owner/name or owner/name@ref), package, to_version (exact or latest), optional from_versionverdict, reason, ordered steps, findings with file:line, unchecked uses with reasons, OSV advisories for both versions, declarations, release, coverage
why_pinnedrepo, packageEvery declaration with file:line, specifier, pins and locked version; for each deliberate pin, introduced_by: the commit and pull request that added it
list_dependenciesrepoDirect dependencies: current version, latest PyPI release, newer_available, and blocked_by (declarations that exclude the latest)
advisoriesrepoOSV.dev advisories for every locked or exactly pinned package, split into affected, not_checked, none_known and unpinned

upgrade_brief reads the current version from a lock file, or else from a single exact == pin. If neither exists, or several versions are found, it asks for from_version.

Example

upgrade_brief with openai/openai-python@v1.54.4, httpx, 0.27.2 → 0.28.0 (abridged):

{
  "verdict": "breaks",
  "reason": "2 change(s) in httpx 0.28.0 break code you use, at 3 place(s)",
  "steps": [
    "In src/openai/_base_client.py:1438, replace proxies= with proxy= or mounts= in the call to `httpx._client.AsyncClient`",
    "In src/openai/_base_client.py:857, replace proxies= with proxy= or mounts= in the call to `httpx._client.Client`",
    "In src/openai/cli/_cli.py:179, replace proxies= with proxy= or mounts= in the call to `httpx._client.Client`",
    "Change `httpx==0.25.2` at requirements.lock:28 to `httpx==0.28.0`"
  ]
}

A first run for a package takes about 10 to 60 seconds; release files are cached afterwards. Long analyses keep the stream open with keepalive pings every 15 seconds.

Next: Connect an agent.