Claude Fable 5 is an extraordinary reasoning model. Paste a diff into chat, ask for a review, and it will sound senior. That is also the trap.
A model that talks like a reviewer is not a PR review agent. Without something that loads the real pull request, scopes the changed lines, runs deterministic checks, and posts inline comments on GitHub (or GitLab / Azure), Fable 5 is just an expensive conversation. It cannot see your CI the way your pipeline does. It cannot prove a finding on path:line. It cannot stop your team from merging “looks fine in chat.”
Kramlipi’s PR review agent is the missing half: customer-hosted tools + scanners + first-pass line comments, with your LLM key—including Fable 5 / Claude when you bring that key. Fable 5 supplies judgment. Kramlipi supplies the review system. Together you get better reviews and you spend fewer tokens doing it.
In this guide:
- Why Fable 5 alone cannot give production-grade code review
- How Kramlipi turns Fable 5 into a real PR reviewer
- Where the cost savings actually come from
- How to try the wedge in about two minutes
Who this is for
- Teams already paying for Claude / Fable 5 who still paste PR diffs into chat
- Eng managers who want first-pass review without another $24–48/seat review SaaS
- Platform and security leads who refuse to upload the monorepo to a multi-tenant review cloud
- Anyone who noticed that eloquent AI comments still miss hardcoded secrets and
eval()
Who this is not for
- Buyers who want an AI to replace senior design review and architecture judgment
- Orgs that need a hosted “upload everything” SaaS and do not care where source lives
- People hunting for a guaranteed fix of every bug—Kramlipi does not claim that
The problem: Fable 5 without a review agent
Fable 5’s strengths—long context, careful reasoning, agentic planning—tempt teams into this workflow:
- Open a PR
- Copy the diff (or “summarize this branch”) into Claude
- Ask for a code review
- Paste comments back by hand—or worse, trust the chat and merge
That workflow fails in four structural ways.
1. No grounded surface
A chat reply is not an inline comment on orders.py:38. Seniors still re-find the line. Juniors still miss the hunk. Review that is not on the PR does not change merge behavior.
2. No deterministic first pass
High-signal issues—hardcoded passwords, dangerous eval, SQL string formatting on changed lines—are cheap for scanners and expensive (and flaky) for a frontier model to rediscover every time. If you only use Fable 5, you pay frontier rates for work a regex/AST pass should catch first.
3. Context waste = cost waste
To “be thorough,” people dump whole files or large ranges into the model. Tokens climb. Latency climbs. Quality does not climb linearly. A PR agent that reviews the diff (plus targeted context) keeps the model on the change set.
4. No honesty bar
Chat can say “LGTM” with confidence. A production review loop should prefer evidence on changed lines, severity labels (must-fix vs should-fix), and quiet when there is nothing actionable. Eloquence is not a merge criterion—same philosophy as verify-gated CI repair.
Bottom line: without Kramlipi (or an equivalent PR agent), Fable 5 cannot give good code review in the sense engineers mean it—actionable, located, repeatable, in the tools the team already uses.
What “good PR review” means here
For Kramlipi, a good first-pass review means:
| Attribute | Chat-only Fable 5 | Fable 5 + Kramlipi PR agent |
|---|---|---|
| Lands on the PR | Manual copy-paste | Inline comments via API |
Tied to file:line | Often vague | Required for findings |
| Secrets / eval / SQL smells | Hit-or-miss, costly | Scanners first, then model |
| Runs where code lives | Your laptop chat | Your runner / laptop / CI |
| LLM key | Anthropic chat session | BYO (Claude / Fable / Gemini / OpenAI / local) |
| Success signal | “Sounds smart” | Actionable findings + human owns design |
We are not claiming the agent replaces staff engineers. We claim it removes repetitive first-pass work so humans spend attention on intent, architecture, and risk acceptance.
How Kramlipi makes Fable 5 better at PR review
Think in layers—not “one prompt to rule them all.”
Layer 1 — Load the real PR
Kramlipi resolves the pull request in your workspace, fetches the diff, and works from changed lines. That alone beats “paste whatever I remembered from the branch.”
Layer 2 — Deterministic scanners and allowlisted linters
Before a frontier model spends tokens, Kramlipi can flag high-signal patterns on the diff (for example hardcoded secrets, dangerous eval, suspicious SQL formatting) and optionally run allowlisted linters on changed lines. Those findings are cheap, repeatable, and easy to explain.
This is the first cost save: do not ask Fable 5 to rediscover what a scanner already knows.
Layer 3 — Model judgment where it earns its keep
Semantic review—subtle logic bugs, missing boundary cases, API misuse—is where Fable 5 shines. Kramlipi calls your configured model with scoped context. You can point CODEAGENTMODEL at a Claude / Fable-class string your LiteLLM setup supports, the same way teams use Gemini or OpenAI today.
Fable 5 is the critic. Kramlipi is the stage manager.
Layer 4 — Publish where review happens
Findings become inline PR comments (and related review surfaces), not a Slack essay. Example receipts on public forks:
Click the discussion anchors. That is the difference between a model and a PR review agent.
Layer 5 — Customer-hosted + BYO key
Source stays on your laptop or your CI runner. The LLM call goes to your provider under your key and DPA. That is the trust story security teams ask for—and it is why “just use Fable 5 in the browser on our private monorepo” does not scale.
Docs: Code review in CI · Get started
Where you save cost (without pretending magic)
Cost is not only “Anthropic invoice.” It is tokens + engineer time + duplicate SaaS seats.
1. Fewer frontier tokens per PR
| Approach | What the model sees | Typical waste |
|---|---|---|
| Chat paste | Oversized dumps, repeated back-and-forth | High |
| Kramlipi review | Diff-scoped context + scanner pre-filter | Lower |
Scanners and linters catch a slice of must-fix issues at near-zero LLM cost. Fable 5 only burns tokens on the residual judgment work.
2. Less human re-work
If the model’s output is already on the PR as line comments, seniors do not re-locate findings. That is usually the larger bill—salary, not API.
3. Avoid paying twice for “review product” seats
Many teams already buy Claude for coding help, then add a per-seat review SaaS. With Kramlipi you keep one reasoning budget (your Fable / Claude / Gemini key) and add an agent that posts reviews on your runner. Team pricing for CI repair sits at a flat org wedge ($49/mo class)—not $24–48 per developer for comments alone. (Check current pricing on kramlipi.com.)
4. No “upload the monorepo” tax
Cloud review products price convenience. You pay with money and with data-boundary complexity. Customer-hosted review keeps the expensive model call narrow and the source local.
Honest caveat: if you point every PR at the largest Fable 5 tier with huge contexts and no scanner pass, you will not save money. The savings assume you use the hybrid loop Kramlipi is built for.
Side-by-side: Fable 5 alone vs Fable 5 + Kramlipi
| Attribute | Chat-only Fable 5 | Fable 5 + Kramlipi PR agent |
|---|---|---|
| Lands on the PR | Manual copy-paste | Inline comments via API |
Tied to file:line | Often vague | Required for findings |
| Secrets / eval / SQL smells | Hit-or-miss, costly | Scanners first, then model |
| Runs where code lives | Your laptop chat | Your runner / laptop / CI |
| LLM key | Anthropic chat session | BYO (Claude / Fable / Gemini / OpenAI / local) |
| Success signal | “Sounds smart” | Actionable findings + human owns design |
How to try it (about two minutes)
Option A — GitHub Actions / CI
Follow the copy-paste workflows:
https://kramlipi.github.io/code-review-ci
Pin example:
- uses: kramlipi/code-agent-action@v0
Set your LLM key as a secret (Gemini, Claude/Fable-compatible via your provider setup, OpenAI, etc.). Review runs on your runner.
Option B — Binary on a laptop
Releases: kramlipi/code-agent-binaries
export GEMINI_API_KEY=your-key # or your Claude / OpenAI env
# point CODE_AGENT_MODEL at your LiteLLM model string
./code-agent experts run code-review --pr 42 -w /path/to/repo
Option C — VS Code
Kramlipi on the VS Code Marketplace
Same idea: agent on your machine, BYO key, review tied to the repo—not a floating chat transcript.
Common mistakes
- Judging review quality by prose. Ask: did it land on the right lines with must-fix evidence?
- Skipping scanners. You will overspend on Fable 5 and under-catch boring vulns.
- Pasting proprietary trees into consumer chat “just this once.” Boundaries erode fast.
- Expecting design-review replacement. First-pass agent ≠ staff architect.
FAQ
Can Fable 5 do code review by itself?
It can discuss code. It cannot, by itself, operate as your GitHub/GitLab PR review agent with inline comments, diff scoping, and CI-native posting. For that you need a review agent—Kramlipi is built for that job.
Does Kramlipi require Fable 5?
No. Kramlipi is provider-agnostic (BYO LLM via LiteLLM-style model strings). Fable 5 is optional horsepower. The point of this article is: if you already invest in Fable 5, pair it with Kramlipi instead of chat-paste review.
Will this cut my Anthropic bill?
It can, when scanners handle cheap findings and the model only sees scoped diffs. It will not if you configure maximum context on every PR. Measure tokens before/after on a sample of PRs.
Does Kramlipi upload my repo to train models?
No. We do not use your repository contents to train models. Core review runs customer-hosted; LLM calls go to your provider with your key.
Where can I see a live example?
Public receipt PRs with inline comments, e.g. kramlipi/skillsaw#1.
Next step
Stop treating Fable 5 as a PR bot. Use it as the brain inside a PR agent that already knows how to load diffs, run scanners, and comment on lines.
Try free: https://kramlipi.github.io/get-started/ CI copy-paste: https://kramlipi.github.io/code-review-ci/ Questions / 15-min live review on your repo: cluevion@gmail.com
Kramlipi — first-pass PR review and verify-gated CI repair on your runner, with your key.