// code intelligence for engineering teams

Understand any codebase in hours, not weeks.

Codewalk indexes a repository, maps its dependencies, answers questions with source citations, and reviews changes with evidence a reviewer can check. Built for the moment you inherit a codebase nobody documented.

Search
cited
Coverage
reported
Keys
BYOK
Tenancy
org-scoped
// search — acme/apiv3 · 412 chunks

where is rate limiting enforced?

confidence0.820retrieval good
app/middleware/rate_limit.py · 41–44
async def enforce(key: str) -> None:    bucket = await redis.hgetall(f"rl:{key}")    if int(bucket.get("count", 0)) > limit:        raise RateLimited(retry_after=bucket["ttl"])
1 of 5 chunks · d=0.1834open file →

// capabilities

Engineered for the questions you actually ask.

// dependency_graph

See how it all connects

Imports are parsed across languages and turned into a dependency map, so an unfamiliar repository has a shape instead of a file tree.

// blast_radius

Know what a change touches

Walk the reverse dependency graph from any file or symbol and get the transitive impact before you start editing, not after review.

// semantic_search

Ask, and get citations

Every answer cites file and line. Each result carries its retrieval confidence, and a weak match is labelled weak rather than dressed up as an answer.

// review_pipeline

Review with evidence

A diff or a ref range is reviewed in batches. Findings come back worst-first with the code that produced them, and a human decides on each one.

// coverage

Honest about what it saw

If a batch fails, the review reports itself as partial and says how much of the diff was never examined. Silent incompleteness is the failure mode this product is built against.

// bring_your_own_key

Your provider, your bill

Codewalk holds no inference account. Chat, reviews and the agent run on the organisation's own provider key, stored encrypted and never readable again.

// how it works

Four steps from a clone URL to a reviewed change.

  1. 01

    Connect a repository

    Point Codewalk at a public clone URL. It belongs to your organisation and nobody else's.

  2. 02

    Index it

    The pipeline clones, parses and embeds the repository. You watch each stage, and a failed run tells you why.

  3. 03

    Ask and search

    Semantic search returns cited chunks with their distance, plus an overview, symbol explanations and a blast-radius map.

  4. 04

    Review and decide

    Send a diff or a ref range. Findings arrive worst-first; accept, dismiss, won't-fix or fix — each decision recorded against the revision it was read from.

// security_model

Your keys. Your providers. Your bill.

Codewalk holds no inference account of its own, and there is no shared pool to fall back on. Every model call runs on a credential the organisation stored, so usage is metered against your own provider agreement.

Provider keys
Encrypted before storage and never returned to a browser again — not even masked. Rotating one replaces the secret in place.
Where code goes
The deployment sends code to the embedding and model providers it is configured with. Those are your accounts; nothing is sent to us.
Machine access
Scoped API tokens for CI and MCP clients, shown once at mint. They cannot administer the organisation or mint credentials.
Tenancy
Everything is scoped to an organisation. Another tenant's repository is not found rather than forbidden.

Start with the repository you are afraid to touch.

Create a workspace, connect a repository, and see what the index knows about it before you add a provider key.