01
Bound the PR
Map every changed range into one complete, evidence-addressable card deck.
Independent Before Consensus
Fresh eyes first.
Fresh eyes first.
Then a verdict you can defend.
Several fresh agents read the complete bounded PR without seeing one another’s answers. Only after every review closes do they challenge candidate issues and vote against explicit evidence thresholds.
Blind first pass
No peer findings or issue counts
Full-PR reads
Every fresh agent gets the same deck
Real quorum
Unique voters; engine-enforced thresholds
reviewer 01
reviewer 03
reviewer 04
reviewer 05
evidence report
diffswarm review engine
Blind discovery
5 fresh reads
claims clustered
evidence survives
Three gates
- Discover independently
- Challenge together
- Vote with evidence
- Enforce quorum
Why It Is Different
+every reviewer sees the complete PR deck
!no conclusion crosses the discovery barrier early
+duplicate identities never inflate consensus
Consensus reached
Confirmed
Featured Workflow
A review surface built for scanning.
Preview the review workflow first, then drill into the pricing and local-first comparison when you need details.
Architecture
Independence is a phase, not a prompt.
Private whole-PR reads converge only after every completion receipt. Discussion and consensus happen on the far side of that barrier.
Pricing
$14.99
Comparison
See the comparison
Unlimited reviews for $14.99/mo.
Bring your own Claude Code or Codex credentials. No token markup, no code proxy.
Local-first instead of per-review cloud review.
Anthropic charges up to $25 per review; DiffSwarm orchestrates on your machine through your existing authenticated model setup.
The Product Contract
No agent influences discovery before the barrier.
02
Discover Blind
Fresh processes review the entire deck independently. Partial exits are discarded, not counted.
03
Deliberate
After the barrier, fresh agents compare anonymized claims and search for support and counterexamples.
04
Enforce Consensus
Unique fresh voters count once. The engine applies configured thresholds and bounds uncertainty.
One CLI. Two Missions.
Choose Your Vector
Bug Hunt Mode
Find Logic & Edge Cases
Track the changed files, follow the live issue list, and inspect exact code or diff context while DiffSwarm checks correctness risks.
$ diffswarm pr <url> --profile balanced
Security Audit Mode
OWASP & CWE Precision
Focus the same repo-first workflow on auth, input handling, config, and crypto changes so security issues stay readable and actionable.
$ diffswarm security <url> --min-severity high
Capabilities
Engineered for CI & Local dev
Bring Your Own Model (BYOK)
DiffSwarm supports Anthropic Claude Code and OpenAI Codex CLI. You provide the credentials and connect directly to that provider. DiffSwarm never proxies inference or marks up model tokens.
CI/CD Integration
Run headless in actions. Use --comment to automatically post verified findings to your GitHub PR.
Auto-tuned Budgets
Pass --time-budget-minutes 60 and the reviewer will scale passes and fleet size to fit the window. Repeated expensive reruns must be explicitly approved.
Local Orchestration, Direct Inference
The terminal app, scheduler, state, and reports run locally. Code, diffs, prompts, and findings are never sent to DiffSwarm servers; selected bounded context goes directly to the Claude or OpenAI provider you configured.
Common Questions
FAQ
What API keys do I need?
You need either Anthropic Claude Code or OpenAI Codex CLI installed and authenticated. DiffSwarm uses your existing credentials to orchestrate the review. No additional API keys from us.
How much do AI tokens cost on top?
Token costs depend on PR size and your AI provider's rates. A typical balanced review uses roughly the same tokens as a normal Claude Code or Codex session. DiffSwarm never marks up or proxies your API calls — you pay your provider directly.
Does my code leave my machine?
DiffSwarm does not receive it. Orchestration and artifacts stay on your machine; the bounded review context is sent directly by your authenticated Claude Code or Codex CLI to your chosen AI provider. DiffSwarm separately calls its API only for subscription entitlement.
How is this different from Anthropic's Code Review?
Anthropic's managed Code Review averages $15–25 per review, is offered to Team and Enterprise plans, and runs on Anthropic infrastructure. DiffSwarm costs $14.99/mo for unlimited orchestration, works with your existing Claude Code or Codex setup, keeps orchestration and artifacts local, and includes a dedicated OWASP security audit mode. Full comparison here.
Can I try before I subscribe?
Yes. Run diffswarm diff sample.diff --dry-run to see the full terminal UI and review flow without a subscription or LLM calls.