Codex MCP server for full codebase context.

Codex starts every task with AGENTS.md and whatever it reads through the shell. On a large repository that means rg, cat and a guess at what an edit reaches, repeated each session. AGENTS.md helps, but it holds only what someone wrote into it.

repowise indexes the repository once, on your machine and with no model calls: the import and call graph, git history (hotspots, co-change, owners, bug-fix history), architectural decisions and per-file health. Codex then asks for what it needs. get_context returns a triage card for a file, get_risk says what an edit reaches and which tests cover it, get_why returns the decision behind a pattern. The managed AGENTS.md section tells Codex which tool fits which question, and a SessionStart hook adds the standing decisions relevant to the session.

Yes. repowise mcp is a stdio MCP server, and repowise init --codex adds it to .codex/config.toml in your repository as [mcp_servers.repowise], with cwd set to the repository and a 20-second startup timeout. codex mcp list shows it. It serves the 10 repowise MCP tools, and repowise mcp --tools with a comma-separated list of tool names narrows the set for a small context window.

The server reads the local .repowise index. Nothing is uploaded, and no tool edits code.

For several repositories, run repowise init . in the parent folder. It finds the git repos up to three levels down, indexes each one, links them through HTTP, gRPC, topic and OpenAPI contracts and cross-repo co-change history, and registers one MCP server that answers for all of them. The workspace server adds list_repos to the 10 MCP tools, and every tool takes a repo argument, or a target written as alias:path. --tools +name adds get_blast_radius, get_architecture and get_conformance for cross-repo impact.

On hosted repowise, one URL answers across a workspace: https://api.repowise.dev/mcp/w/{workspace_id}, with the same API key as the per-repo URL. That shape needs Pro, Teams or Admin.

On Codex, against a bare agent, repowise cut output tokens by 31.6% and reached an answer in 3.8 tool calls instead of 7.2 (n = 43 · p = <0.0001 · Codex (gpt-5.6-sol) · django/django at a single pinned commit · 2026-08). On Claude Code the same design cut output tokens by 15.9% (n = 15 · p = 0.035 · Claude Code (claude-sonnet-5) · django/django at a single pinned commit · 2026-08).

The saving comes from fewer reads: a triage card instead of a whole file, one symbol's lines instead of the module, one cited answer instead of a chain of searches.

repowise distill, measured with the CLI one command at a time: 89% fewer tokens on git log -50 and 61% on a pytest run with 11 failures, with all 11 failure lines kept.

Inside Codex, get_change_risk scores uncommitted work or a base..HEAD range against the repository's own commit history: defect risk, files that usually change with these but are missing, and the tests that reach the diff. /prompts:repowise-risk runs it as a slash prompt. That is the check to run before you ask Codex's /review, or before you push.

The repowise PR bot is a GitHub App. On a pull request it stays silent unless a signal fires: files that usually change together but are missing from the diff, tests that import the changed files, a health drop. Its verdict is the Repowise review check run. The optional AI review on Pro attaches a prompt for your agent to each finding, and the pull request page on repowise.dev has a Copy fix prompt for your agent button. On GitLab, Azure DevOps and Bitbucket there is no bot; indexing and the CI gates (patch coverage, doc drift, new secrets, change risk) still work.

To fix a finding, copy its prompt and paste it into Codex. The prompt names the files and the finding, so Codex starts from the evidence instead of re-reading the pull request.

Hosted repowise serves the same tools over Streamable HTTP at https://api.repowise.dev/mcp/{owner}/{repo}, with an API key from Settings, Coding agents (MCP). The per-repo URL works on every plan, Free included (10 answers a month on Free). Claude on the web and desktop connects with a custom connector and sign-in, and ChatGPT takes the same MCP address. Codex can use the hosted URL wherever its MCP config accepts a Streamable HTTP server with an Authorization header, and the local stdio server above is the default route.