How Tree-sitter concrete syntax trees, ephemeral RAM execution, and sub-second multi-model failover power zero-latency code reviews.

The Problem with Raw LLMs

Traditional AI review tools feed raw 2,000-line git diffs directly into commercial LLM context windows. This approach causes:

The 5-Stage Pipeline

GitCaptain addresses this with a compiler-inspired architecture:

[GitHub Webhook]
       │ (HMAC Verified in 2ms)
       ▼
[Tree-sitter CST Engine] ─── Isolates modified AST function scopes in RAM
       │
       ▼
[Targeted Prompt Assembly] ─── 70% fewer tokens, includes exact parent symbol signatures
       │
       ▼
[BYOK Inference Router] ─── Sub-500ms failover across Google / Groq / OpenRouter
       │
       ▼
[GitHub Markdown Emitter] ─── Inline review comments & complexity radar metrics
       │
       ▼
[tmpfs Zero-Fill Wipe] ─── Complete in-memory teardown. Zero code on disk.

Tree-sitter AST Delta Isolation

When a PR webhook arrives, GitCaptain's worker loads the modified files into a temporary in-memory Linux virtual filesystem (tmpfs).

Sub-Second Multi-Model Failover

If your primary model provider (e.g., Google AI Studio) responds with a 429 Rate Limit or temporary outage, GitCaptain's worker automatically reroutes the prompt payload to your secondary configured fallback provider (e.g., Groq Llama-3.3-70B) in under 500ms. The PR webhook never drops.

Ephemeral RAM Teardown

Once review comments are delivered to GitHub's REST API, the worker unmounts the tmpfs partition and zeros the memory space. No user repository source code is ever written to disk, and no code is ever used for model training.

Connect your personal API keys directly from Google AI Studio, Groq, Mistral, or OpenRouter for 100% free individual usage with zero token markups.

Why Bring-Your-Own-Keys?

Most AI developer tools act as billing middlemen, charging $20-$40/seat/month while marking up token prices by 300-500%.

GitCaptain is fundamentally different:

Supported Inference Providers

Google AI Studio (Recommended for $0 Free Tier)

Groq Cloud (Recommended for Speed)

OpenRouter & Custom Gateways

Key Storage & Cryptographic Security

Your API keys are encrypted at rest using AES-256-GCM with a unique Initialization Vector (IV) per record.

Understand the scope, architectural blast radius, and exact line-level improvements of every pull request.

PR Walkthrough Comments

On every pull request, GitCaptain generates a concise, human-readable walkthrough comment pinned to the discussion thread:

Complexity Radar & AST Metrics

GitCaptain analyzes AST node depth and cyclomatic branches to compute code complexity metrics:

1-Click GitHub Suggestions

When GitCaptain identifies an optimization or bug, it formats inline comments as GitHub Suggestion diff blocks:

- const user = await db.users.find({ id: req.params.id });
- if (!user) return res.send(404);
+ const user = await db.users.findById(req.params.id).lean().exec();
+ if (!user) {
+   return res.status(404).json({ error: 'User not found' });
+ }

Engineers can click Commit suggestion directly in GitHub's UI to apply the fix in one click.

Interactive Comment Commands

You can interact with GitCaptain directly by mentioning @gitcaptain in any PR comment:

Learn why GitCaptain never stores your source code on disk or uses your intellectual property for model training.

Zero Code Retention Pledge

At GitCaptain, we believe your codebase is your most valuable intellectual property. Our architecture is built from the ground up around Zero Data Retention (ZDR):

Temporary tmpfs Sandboxes

When a review job starts:

# Internal worker memory mounting
mount -t tmpfs -o size=256M,noexec,nosuid,nodev tmpfs /sandbox/pr-1284

The worker checks out the git diff into this isolated sandbox, runs Tree-sitter parsers, and dispatches the prompt. Upon task completion, the mount is unmounted and freed:

umount /sandbox/pr-1284

No Public Model Training

Under our Bring-Your-Own-Key model: