repobrain vs. repowise

A head-to-head comparison of every capability, architecture decision, and tradeoff.

---

At a Glance

Metric	repowise	repobrain
Initial indexing (1,000 files)	~25 min	~3–5 min
Git history depth	500 commits (hardcoded)	10,000 (configurable)
Missing import edges	20–40%	~0% (dynamic hints)
Silent consistency failures	5–15%	0% (atomic transactions)
MCP tools	8	12
CLI commands	4	6
Temporal scoring	❌	✅ exponential decay
PR blast radius	❌	✅
LLM cost tracking	❌	✅
License	AGPL-3.0	MIT
Self-hosted	✅	✅
RAG used during generation	❌	✅

---

Full Feature Matrix

Documentation Generation

Feature	repowise	repobrain
Generates file-level docs	✅	✅
Generates symbol-level docs	✅	✅
Fetches dependency context before generation	❌	✅
Includes centrality score in prompt	❌	✅
Includes hotspot score in prompt	❌	✅
Freshness / staleness tracking	Basic	Advanced
Semantic search over generated docs	✅	✅

Git Analysis

Feature	repowise	repobrain
Commit history depth	500 (hardcoded)	10,000 (configurable)
Hotspot scoring	Linear count	Exponential decay
Co-change analysis	✅	✅ temporal-weighted
Ownership calculation	✅	✅ temporal-weighted
Architectural decision capture	✅	✅
Percentile rank refresh on incremental update	❌	✅

Graph Construction

Feature	repowise	repobrain
Python parsing	✅ tree-sitter	✅ tree-sitter
TypeScript parsing	✅	✅
Go parsing	✅	✅
JavaScript parsing	✅	✅
Django dynamic imports	❌	✅
pytest fixture imports	❌	✅
Node.js require() patterns	❌	✅
PageRank centrality	✅	✅
Transitive dependency traversal	✅	✅

Storage & Reliability

Feature	repowise	repobrain
SQL store (SQLite)	✅	✅
Vector store (LanceDB)	✅	✅
Graph store (NetworkX)	✅	✅
Atomic transactions across all 3	❌	✅
Rollback on partial failure	❌	✅
Consistency failure rate	5–15%	~0%

MCP Tools

Tool	repowise	repobrain	Improvements
explain_file	✅	✅	RAG-injected dependency context
explain_symbol	✅	✅	Same
get_hotspots	✅	✅	Temporal decay scoring
get_ownership	✅	✅	Temporal-weighted
get_dependencies	✅	✅	Dynamic hint edges included
get_architectural_decisions	✅	✅	Same
search_codebase	✅	✅	Same
get_cochange_patterns	✅	✅	Temporal-weighted
get_pr_impact	❌	✅ NEW	—
get_knowledge_map	❌	✅ NEW	—
get_test_gaps	❌	✅ NEW	—
get_security_hotspots	❌	✅ NEW	—

CLI

Command	repowise	repobrain
index / init	✅	✅ faster
serve	✅	✅
query	✅	✅
status	✅	✅
review <PR>	❌	✅ NEW
costs	❌	✅ NEW

Security & Privacy

Feature	repowise	repobrain
Self-hosted	✅	✅
Code stays local	✅	✅
BYOK (bring your own API key)	✅	✅
No telemetry	✅	✅
GitHub webhook HMAC-SHA256 validation	Partial	Full
License	AGPL-3.0	MIT
Commercial use without license fee	❌ (AGPL)	✅ (MIT)

---

Architecture Comparison

Documentation Generation — The #1 Flaw

repowise populates a vector store during indexing but never queries it when generating documentation. Each file's docs are generated with no knowledge of its dependencies.

# repowise (simplified)
async def generate_docs(file):
    prompt = f"Document this file:\n{file.content}"
    return await llm.complete(prompt)  # no dependency context

repobrain fetches dependency documentation from LanceDB before every LLM call:

# repobrain
async def generate(file_path, parse_result, graph):
    dep_paths = graph.get_direct_dependencies(file_path)
    dep_contexts = await retriever.get_docs(dep_paths)  # ← the fix
    prompt = build_prompt(
        file_content=parse_result.content,
        dependency_contexts=dep_contexts,   # actual dep docs included
        centrality=graph.get_centrality(file_path),
        hotspot_score=metrics.get_score(file_path),
    )
    return await llm.complete(prompt)

Storage Consistency — Silent Failures

repowise writes to three independent stores with no coordination. Any write can fail silently, leaving indexes in an inconsistent state (5–15% failure rate under load).

# repowise — three independent writes
await sql_store.save(file_record)
await vector_store.upsert(embedding)
graph_store.add_node(file_path)   # if this crashes, SQL and vector are already written

repobrain wraps all three in an atomic transaction — all succeed or all roll back:

# repobrain — atomic across all 3 stores
async with coordinator.transaction() as txn:
    txn.pending_sql_calls.append(lambda: sql.save(file_record))
    txn.pending_vector_records.append(embedding)
    txn.pending_nodes.append(file_path)
# On any exception: SQL rollback + vector delete + graph node removal

Hotspot Scoring — Temporal Blindness

repowise treats a commit from 3 years ago the same as a commit from yesterday:

# repowise — linear accumulation
score = sum(commit.lines_changed for commit in commits[-500:])

repobrain applies exponential decay so recent activity dominates:

# repobrain — temporal decay
from math import exp, log

for commit in commits:
    age_days = (now - commit.authored_date).days
    decay = exp(-log(2) * age_days / halflife_days)   # halflife default: 180 days
    normalized = min(commit.lines_changed / 100, 3.0)  # cap outliers
    score += decay * normalized

---

Performance Comparison

Benchmark	repowise	repobrain	Improvement
1,000-file Python repo — first index	~25 min	~3–5 min	5–8× faster
1,000-file repo — incremental update (50 files)	~3 min	~30 sec	6× faster
Embedding concurrency	Sequential	ThreadPoolExecutor + semaphore	parallel
Parse stage	Single process	ProcessPoolExecutor (multi-core)	CPU-bound speedup
Git + parse stages	Sequential	asyncio.gather (concurrent)	concurrent

---

Licensing

Aspect	repowise	repobrain
License	AGPL-3.0	MIT
Commercial use	Requires commercial license	Free — no restrictions
Modify and redistribute	Must open-source changes	No requirement
SaaS / hosted deployment	Requires commercial license	Free
Cost	Free (open-source) + paid plans	Free — MIT forever

repowise is AGPL-3.0 — if you deploy it in a commercial product or SaaS, the AGPL requires you to release your entire application under AGPL or purchase a commercial license from repowise-dev.

repobrain is MIT — use it commercially, integrate it into proprietary software, deploy it as a SaaS. No license fees, no restrictions.

---

Migration from repowise

repobrain is API-compatible with repowise. All 8 repowise MCP tool names work unchanged. Migration takes under 5 minutes.

See the Migration Guide →