Feed a Public Repo to Claude Code Without Burning Context: Repowise's Free MCP Setup and Its Four Limits

When Claude Code or Cursor works on an unfamiliar codebase, the expensive part is rarely the reasoning. It is the agent grepping, opening and re-opening files to reconstruct context it should have been handed.

Repowise’s answer: index the repo once and serve it over MCP. The published benchmark figures in its official guide:

• 2,391 tokens to load context instead of 64,039 — 96% fewer, roughly 27×
• 89% fewer file reads across paired benchmarks at answer parity
• 70% fewer tool calls on the flask48 and sklearn48 suites
• Ten task-shaped MCP tools over one endpoint, open source under AGPL-3.0

And for public repos, the free tier covers it. Here is the flow and where free stops. Verified 2026-09-07.

1. The problem it targets

The guide states it plainly: entity-by-entity tools force the model to do retrieval itself — search, open, scroll, repeat. That produces four costs:

1. Fragmented context: the agent sees file fragments and never the whole picture.
2. Compounding token cost: raw exploration of a single task can spend 64,039 tokens just loading context, before any reasoning.
3. Retries multiply reads: a missed grep means another search, another open, another scroll.
4. No enrichment: raw file reads carry no ownership, no history, no risk and no “why”.

2. Four steps, per the official pipeline

The fourth step is the one worth copying: it does not pretend the index is always right; it makes freshness part of the response and lets the agent judge. The classic failure mode for cached context is treating stale data as truth, and this design makes that risk visible.

Endpoint registration differs by client (Claude Code, Cursor, Cline and Codex each configure MCP in their own place) and the official docs cover each; the concept is that one URL is all you register.

3. The four free-tier ceilings

If you plan to run public repos on free, know these (all listed on the pricing page):

1. Public repo indexing up to 250 MB. Above that you need Pro (5 GB).
2. 500 commit analysis limit. Free excludes full commit history, so the git intelligence — hotspots, ownership, bus factor — is built from the most recent 500 commits.
3. 10 AI answers per month. This is the hard one, though deterministic features (health score, dependency graph, hotspots, dead code, ownership) never consume it — only real AI answers do.
4. No push-triggered auto-sync. Free refreshes weekly, with the health badge updating alongside, rather than rebuilding on every push.

One more that surprises people: repos below 1,000 stars get no written explanations on free. Your own small project gets the index, graphs and search, but not the auto-written prose.

4. Deciding whether to upgrade

Three questions, no guessing required:

• Will you exceed 10 AI answers a month? If so, Pro’s US$15 includes US$5 of model credit, making the software itself effectively US$10.
• Do you need private repos? Free is public-only; private indexing via the GitHub App is Pro.
• Do you need updates immediately after a push? Weekly is fine for a stable project and not fine for a fast-moving one.

Full plan breakdown: US$5 of Repowise’s US$15 plan is model credit.

5. Two caveats

1. Saving context tokens is not saving everything. The 96% applies to loading context; the reasoning still costs what it costs. The real payoff shows up in long sessions where the agent no longer re-greps each time.
2. Source is processed transiently and not persisted (as stated), and the product is self-hostable. But if your organisation has a written rule against source passing through a third-party service, the self-hosted AGPL-3.0 route is the compliant one, not the hosted tier.

Sources: Repowise’s AI Context & MCP guide (by Raghav Chamadiya, marked updated July 2026) and its pricing page, read directly on 2026-09-07. Every benchmark figure (96%, 89%, 70%, 2,391 vs 64,039) is the vendor’s own published paired-run result; we did not reproduce them. Plan summary also on our Repowise tool page.