v3.1.0 Git-Aware Proxy • Zero Cloud Roundtrips

Stop paying your coding agent to
re-read the same files.

Local proxy that replays unchanged repository reads in 0.08ms—cutting token burn by 79%.

Execution Time

6,130ms → 1,450ms

76.3% faster

Prompt Token Burn

3,280 → 680 tokens

79.3% saved

STEP 1 Install & launch

❯pip install omnicache-proxy && omnicache run claude

STEP 2 Configure Claude Code

BASH / ZSH ENVIRONMENT

❯export ANTHROPIC_BASE_URL="http://127.0.0.1:8000" && claude

Docker image: ghcr.io/13manmayarai-hash/omnicache:latest [TODO: confirm published image]

Architecture ↓

Claude Code CLI Compatible

Cursor IDE Compatible

Windsurf Agent Compatible

Aider Pair Programmer Compatible

Anthropic Claude 3.5 Supported

OpenAI GPT-4o / o1 Supported

Claude Code CLI Compatible

Cursor IDE Compatible

Windsurf Agent Compatible

Aider Pair Programmer Compatible

Anthropic Claude 3.5 Supported

OpenAI GPT-4o / o1 Supported

read_file("package.json") 0.08ms HIT

list_dir("tests/") 0.06ms HIT

git status subpath 0.04ms HIT

pytest inspection 0.11ms HIT

npm test runner 0.09ms HIT

cargo clippy check 0.08ms HIT

read_file("package.json") 0.08ms HIT

list_dir("tests/") 0.06ms HIT

git status subpath 0.04ms HIT

pytest inspection 0.11ms HIT

npm test runner 0.09ms HIT

cargo clippy check 0.08ms HIT

BENCHMARK TRACE

Why Coding Agents Bleed Tokens

Execution trace across 4 consecutive read operations on an unchanged repo.

Vanilla Agent Mode Cloud Egress Every Turn

Turn 1: read_file("src/config.py")

Remote Egress 1,450 ms • 680 tok

Turn 2: re-read("src/config.py")

Redundant 1,380 ms • 680 tok

Turn 3: list_dir("tests/")

Redundant 1,890 ms • 1,240 tok

Turn 4: re-read("src/config.py")

Redundant 1,410 ms • 680 tok

Total Latency

6,130 ms

Total Tokens

3,280 tokens

With OmniCache Sidecar 100% Localhost Intercept

Turn 1: read_file("src/config.py")

Cold Miss 1,450 ms

Turn 2: re-read("src/config.py")

Git Match 0.08 ms • 0 tok

Turn 3: list_dir("tests/")

mtime Match 0.06 ms • 0 tok

Turn 4: re-read("src/config.py")

Git Match 0.08 ms • 0 tok

Total Latency

1,450.2 ms (-76%)

Tokens Saved

2,600 tokens ($0.0078)*

*Claude 3.5 Sonnet @ $3/1M tokens (2,600 of 3,280 prompt tokens eliminated).

99.4%

Replay Speedup

1,450ms remote → 0.08ms RAM

79.3%

Token Reduction

2,600 / 3,280 repeat tokens cut

0.08 ms

P50 Latency

Local SQLite WAL index lookup

CACHE INTEGRITY

Deterministic Freshness Guarantees

Cryptographically bound to local Git tree deltas and filesystem mtimes.

01 • FILE MTIME

Nanosecond Timestamps

Verifies st_mtime_ns and byte size on every read. Local editor saves evict the cache key instantly.

02 • TREE DELTA

Git Index Probing

Probes .git/index and untracked status. Staged or unstaged workspace diffs force fresh evaluations.

03 • WRITE-THROUGH

Mutation Invalidation

Mutations (write_to_file, bash) immediately evict affected paths from SQLite WAL memory. [TODO: confirm fallback scope for non-git workspaces]

ARCHITECTURE MAP

From Tool Call to Zero-Token Replay

Local execution pipeline from agent HTTP request to RAM replay.

STEP 01

Protocol Intercept

Catches localhost HTTP payloads; coalesces concurrent calls via SingleFlight.

SingleFlight Active • 0ms Delay

01 • Local Intercept

Merkle Guard • 02

STEP 02

Git SHA Verifier

Verifies working-tree Merkle root and mtimes; evicts stale keys before replay.

0.04ms Check • Merkle Clean

STEP 03

L1 Memory Replay

Direct SQLite WAL lookup; replays tool response to agent in 0.08ms.

0.08ms P50 • RAM Throughput

03 • L1 Memory Replay

Int8 Semantic L2 • 04

STEP 04

Int8 Semantic L2

Int8 CPU cosine vectors evaluate semantic hits in 1.02ms; forwards misses upstream.

1.02ms Match • 0 Cloud Leak

Instant Tool Replay Delivered

0.08ms P50 Latency • 0 Cloud Tokens • $0.00 Cost

AUTO-PROBE INTERCEPT EXECUTED HTTP/1.1 200 OK

INTERCEPTED: read_file("package.json")

Replay Latency

0.08 ms (vs 1,450ms)

Tokens Saved

680 tokens ($0.0020)

Workspace Guard

HEAD 80e1080 • Clean

Zero Hallucination

Git SHAs & nanosecond mtimes evict stale code instantly.

0.08ms P50

In-memory SQLite WAL eliminates remote latency.

Air-Gapped

Zero cloud telemetry, code diffs, or credentials leave localhost.

79% Token Cut

Eliminates redundant prompt tokens across multi-turn loops.

GATE EVALUATION

Deterministic Replay Sandbox

Interactive 5-gate pipeline evaluation on localhost.

Example Traces:

GATE 01 INTERCEPT

GATE 02 GIT DELTA

GATE 03 RADIX L1

GATE 04 INT8 L2

GATE 05 WAL REPLAY

SECURITY SPECIFICATION

Local-First Privacy Architecture

Air-gapped daemon on 127.0.0.1:8000. No code, keys, or telemetry leave your machine.

01 • INGRESS

Local Intercept

Evaluates tool payloads locally on localhost HTTP. Forwarded headers pass directly to official provider endpoints. [TODO: confirm header capture scope]

02 • STORAGE

On-Disk SQLite WAL

Tool responses and Merkle hashes persist to ~/.omnicache/cache.db in WAL mode. Zero cloud sync. [TODO: confirm default DB path]

03 • EGRESS

Direct Provider TLS

Cache misses stream directly to official provider TLS endpoints (e.g. api.anthropic.com). Zero telemetry. [TODO: confirm telemetry verification]

LIFECYCLE • RESET

Instant Purge & Bypass

Run omnicache cache clear or delete ~/.omnicache/cache.db. Run unset ANTHROPIC_BASE_URL to bypass instantly. [TODO: confirm CLI clear flag]

ECOSYSTEM COMPATIBILITY

Works With Your Agents & Tools

Transparent drop-in for Anthropic and OpenAI HTTP tool-calling protocols.

Supported Agents

Claude Code CLI Native Drop-in
Cursor AI Agent Base URL Support
Windsurf IDE Localhost Proxy
Aider Pair Programmer CLI Base Flag
Custom Agent Loops HTTP Protocol

Target http://127.0.0.1:8000 as base URL. [TODO: confirm extended agent matrix]

Supported LLM Providers

Anthropic Claude 3.5 & Opus Direct TLS Pass
OpenAI GPT-4o & o1 Proxy Schema Pass
Local Models (Ollama, vLLM) OpenAI Endpoint
SingleFlight Coalescing Active In-Memory
Model Context Protocol (MCP) stdio & SSE Mode

Standard provider protocol pass-through. [TODO: confirm OpenAI provider test suite]

FREQUENTLY ASKED QUESTIONS

Technical Operations

What happens when a file changes on disk?

Nanosecond st_mtime_ns checks trigger immediate cache eviction. Never serves stale code.

Does OmniCache work completely offline?

Cached tool calls replay 100% offline from local SQLite WAL memory. Uncached queries route upstream. [TODO: confirm offline scope]

Where are API credentials stored?

Never stored. ANTHROPIC_API_KEY passes through in-memory TLS directly upstream. [TODO: confirm credential security]

How do I bypass or uninstall OmniCache?

Run unset ANTHROPIC_BASE_URL to bypass immediately, or pip uninstall omnicache-proxy to remove.

Which coding agents are supported out of the box?

Claude Code CLI native; any agent supporting custom HTTP proxy URLs (Cursor, Windsurf, Aider). [TODO: confirm matrix]

How is cache storage managed on disk over time?

SQLite WAL mode with automatic LRU eviction bounds disk usage. [TODO: confirm default storage cap]

ZERO CONFIGURATION Instant localhost setup

Deploy Local Memory Replay in 30 Seconds

Replay repetitive tool calls at 0.08ms with zero cloud token burn.