Blind grep loops
Without an index, an agent guesses a keyword, greps, misses, and guesses again. Every miss is another round-trip before the real work starts.
Self-hosted · Runs air-gapped · No telemetry
Codebase intelligence
Codebase intelligence
you own and control.
For teams and agents moving through legacy or complex systems. CodeNavi runs on your hardware and gives Claude Code, Antigravity, or any MCP client ranked, bounded answers about your codebase — ~80% fewer context tokens per task — instead of letting it grep blindly. Your code never leaves your machine, and the index is an asset you keep, not a subscription you rent.
Request takes a minute — your signed license arrives by email within 24 hours.
Agents spend their context window finding code, not fixing it
Whole files for one function
The relevant code is 20 lines; the agent reads the whole 2,000-line file — and its neighbors — to be sure. That's the context window gone by mid-task.
Confident answers, no receipts
Ask “what breaks if I change this?” and a model without the call graph gives you a plausible guess. Plausible is not what you refactor on.
CodeNavi replaces that loop with ranked retrieval — RAG purpose-built for code: the few chunks that matter, cited to file:line, under a token budget — ~80% fewer context tokens per task.
Smarter agents, grounded in your real code
More of the context window left for thinking
The agent reads the few chunks that matter, not whole files: ~80% fewer context tokens per task. The space it saves goes to reasoning about your problem. You also get fewer compactions mid-task and a lower API bill.
Fewer hallucinations, answers with receipts
The agent works from code it actually found in your repo, not from what similar projects usually look like. So it is much less likely to invent a function, a parameter, or a file path. Every result is cited to file:line, so you can check each claim.
Check the blast radius before the edit
Before a change, find_callers and impact_analysis show what depends on the code. After it, diff_impact reviews the diff. Hub functions are flagged, so the agent knows when to slow down.
An index that keeps up with the agent
The agent can reindex its own changes in about 5 seconds, and read_file always reads the live disk. Search results flag a stale index, so the agent never edits from an old picture of the code.
Memory that outlives the session
Agents save what they learn as notes tied to your code. The next session, or a different agent, finds them in search. A note is marked stale when the code under it changes.
Safe to give to an autonomous agent
Secrets and PII are redacted before they reach the model. Prompt injections hidden in code comments and docs are flagged. CodeNavi only reads your repos and never writes to them. How it works →
Sits inside your environment. Only answers come out.
People
EngineersNew developersDevSecOpsComplianceExecutives
AI Agents
Claude CodeCodexAntigravity10+ othersLocal models
QuestionsSymbol lookupsFile readsDiffs to reviewImpact checks33 MCP tools ↗
Ranked answersFile:line receiptsSecrets stripped
Environment you own and control
reads & indexes — read-only
CodeNaviLicensed
- Finds code by meaning, not string matches
- Traces what a change touches: callers, impact, diffs
- Strips secrets and injected prompts before they leave
- Parses 15 languages structurally, legacy Perl included
- Compiles your docs into a searchable wiki
- Keeps the index on your GPU — an asset you own
Your infrastructure
100% on-prem / VPCYour reposYour docsYour GPUYour network
That is everything that crosses the wall. No telemetry, no cloud index, works air-gapped.
Everything in the box
Self-hosted — your code stays with you
Self-hosted container — Docker or Podman (rootless works, so Fedora/RHEL shops that don't allow Docker are covered) — in your VPC. Embeddings, crawl history, and chat logs never leave disk you own. Zero telemetry, zero SaaS dependency. Chat and wiki default to local models (Ollama, vLLM). Runs on your NVIDIA or AMD fleet — including Instinct MI300 — fully local.
A code-native embedding model
A code-specialist embedding model ranks structurally similar functions. Vector + AST + grep retrieval — not autocomplete, not blind keyword search.
15 programming languages
PHP, JS, TS, Python, Go, Java, Ruby, Rust, C#, Bash, Fortran, Perl/CGI, Swift, Kotlin, Scala. AST chunks respect function and class boundaries. Dedicated chunking for SQL, Smarty, and Template Toolkit files.
33 tools over one protocol
33 tools (41 with the optional crawler) exposed over the Model Context Protocol. Drop into any MCP-compatible client via one config. Mirror CLI for CI scripts and ops automation. Browse every tool →
Ranked, bounded answers under a token budget
Top-N chunks with similarity scores and file:line citations. The agent gets a ranked, decision-ready list — not a haystack of files to read.
“What breaks if I change this?” — answered from a call graph
find_callers, impact_analysis, and diff_impact trace a change through a call graph built at index time — high-traffic hub functions flagged, every hop cited. Receipts, not a plausible guess.
Post-deploy web monitoring — optional crawler
Flip on the crawler and CodeNavi watches your running site too: compare_runs diffs two crawls to surface every new error and broken page since your last clean deploy, search_errors makes your error history semantically searchable, and get_referrers maps which pages link a URL before you change it.
Multi-codebase, one server
Scope queries to one repo or search across all of them. Index checkout, tracking, admin, and platform code in a single install — scoped MCP endpoints per repo.
Incremental reindex in seconds
Delta reindex of changed files only — about 5 seconds for a typical push. Trigger it from a git hook or your MCP client; no nightly batch jobs, no stale results.
Secrets and PII redacted at chunk time
API keys, JWTs, cloud credentials, DSNs, and private keys are matched and rewritten to [REDACTED:rule_id] before they reach any model, index, or log. PII is covered by default too: payment card numbers (Luhn-validated) and IBANs (mod-97-validated), with an opt-in extended set. See how security works →
OWASP gap analysis, evidence included
Run /codenavi:compliance from the bundled Claude Code plugin and get a gap report against OWASP ASVS 5.0 core or the OWASP Top 10 2025 — one verdict per rule, each backed by a file:line citation from your indexed code. Audit preparation you can hand to a security reviewer — not a certification.
Free 30-day trial, then a simple license
Evaluate the full product free for 30 days on your own infrastructure, then continue with a straightforward commercial license. Self-hosted, no telemetry, no per-seat metering — a signed image you run, not a SaaS you depend on.
Secrets and PII redacted, injections flagged — before anything reaches a model
Every chunk is scanned on hardware you own. See the layers, the rules, and how to verify it yourself.
Your agent, answering from the index
Real footage — an agent asking CodeNavi plain-English questions and getting ranked, cited answers grounded in real code. Every result is a real file:line.
One index. Every agent you use.
Claude Code, Claude Desktop, OpenAI Codex, Google Antigravity, VS Code, Zed, OpenCode, Pi, AnythingLLM — anything that speaks MCP queries the same local index, CI review bots on GitHub or Forgejo Actions included. Pick your client and get the exact config, ready to paste. Claude Code, Codex, and Antigravity also get bundled plugins — slash commands, guided setup, and a statusline badge on top of the raw MCP connection.
Common questions
How is this different from Cursor's indexing?
Cursor indexes your code on its servers for in-IDE chat. CodeNavi indexes locally and exposes the index as an MCP server, so Claude Code, Cursor, Codex, or your own agent can all query the same local index without sending source to a vendor. The crawler half — post-deploy regression detection via compare_runs, search_errors, get_referrers — is unique to CodeNavi.
Do I need a GPU?
For indexing, yes — an NVIDIA GPU, RTX 20-series or newer (Linux, or Windows 11 via Docker Desktop; ~11 GB VRAM handles small-to-mid repos, 16 GB+ is comfortable for large codebases; pre-RTX / GTX 10-series cards aren't supported), or an AMD GPU on Linux (Radeon RX 6000-series or newer, or Instinct MI100/MI200/MI300) via the ROCm build of the image — same index quality, verified vector-for-vector against the NVIDIA build. Or pre-build the index on a cloud GPU and ship the result to a CPU-only host. Once indexed, CodeNavi serves queries at full quality on any machine, including Apple Silicon Macs under Docker.
Does my code leave my machine?
No. Embeddings, indexing, search, and reranking all run locally — and the LLM features (code chat, docs wiki) default to local models too, via Ollama, vLLM, or any OpenAI-compatible endpoint. Paired with a local agent, the whole loop runs air-gapped. Cloud LLMs (Claude, OpenAI, OpenRouter) are the opt-in, not the default — and even then receive ranked chunks, not whole files. Zero telemetry from CodeNavi itself.
Can it run fully air-gapped?
Yes. The image ships with the embedder, reranker, and every dependency baked in — nothing resolves or downloads at runtime — and the installer has an explicit air-gapped mode. Pair it with a local LLM (Ollama, vLLM) and a local agent and the entire RAG loop runs inside your network boundary. The optional once-a-day update check is disabled by deleting one file.
Which languages does it understand?
Fifteen languages get full structural parsing — PHP, JavaScript/TypeScript, Python, Go, Java, Ruby, Rust, C#, Bash, Fortran, Swift, Kotlin, Scala, and Perl/CGI — plus SQL, templates, config, and docs formats. Legacy codebases are a first-class target, not an afterthought. Full matrix in the docs.
Which AI tools does it plug into?
Claude Code, Claude Desktop, Google Antigravity, OpenAI Codex, VS Code, Zed, OpenCode, Pi, AnythingLLM, and CI review bots on GitHub or Forgejo Actions — anything that speaks the Model Context Protocol. One index serves them all, and a guided connect wizard writes the config for you. Step-by-step guides per tool.
What does it cost?
Start with a free 30-day trial — self-hosted, no account needed, governed by the CodeNavi Evaluation License. After that it's a simple commercial license: a signed image you run on your own infrastructure, not a SaaS subscription. Pricing is simple and self-serve — not enterprise SaaS contracts — contact us for a license or pricing chat.
30-day free trial
Try CodeNavi for 30 days — free.
Signed 30-day trial license, self-hosted on your infrastructure. No telemetry. No account needed. Receive your license by email within 24 hours.
Not ready to install?
New releases, product updates, and the occasional tip. No noise.
By subscribing, you agree to receive product updates. We don't sell your data. See our privacy policy.