bibividi 0.2.0

Token optimization middleware for AI APIs

Bibividi

Token/Quota optimization middleware for AIs Intercepts Anthropic / OpenAI / Gemini / Groq / Mistral / Ollama requests, runs a matching Python script locally, and sends a condensed prompt — reducing token consumption by 60–95%. Turn each solved prompt into a reusable program. Your library grows. Your AI bill shrinks — or your free-tier daily quota goes further.

Your app/Your prompt ──► BIBIVIDI proxy :8080 ──► Anthropic / OpenAI / Gemini
                   │
         Matching engine (embeddings)
                   │
     Script found → local Python execution
                   │
         Condensed context (90% fewer tokens)

No prompt is ever stored. Processing is 100% local.

---

Quick start

Desktop app (recommended) — download the installer for your platform from Releases. The proxy starts automatically on launch.

CLI

pip install bibividi
bibividi start                    # proxy starts on :8080
bibividi add scripts/pdf-clause-extractor/
bibividi status

Integrate in 2 lines

Anthropic SDK

import anthropic
from bvd.sdk import anthropic_http_client

client = anthropic.Anthropic(
    api_key="sk-ant-...",
    http_client=anthropic_http_client(),   # ← only change
)
# All calls now route through BIBIVIDI — zero code changes elsewhere

OpenAI SDK

import openai
from bvd.sdk import openai_http_client

client = openai.OpenAI(
    api_key="sk-...",
    http_client=openai_http_client(),      # ← only change
)

Async variants

from bvd.sdk import async_anthropic_http_client, async_openai_http_client

client = anthropic.AsyncAnthropic(http_client=async_anthropic_http_client(), ...)

---

What is a BSS skill?

A BSS (BIBIVIDI Skill Specification) pack is a directory with two files:

File	Role
skill.json	Metadata: id, name, description, tags, hash
skill.py	Python script. Reads PROMPT, prints condensed output to stdout

# skill.py — PROMPT is injected automatically
import re, json

clauses = re.findall(r'(?:Article|Clause)\s+\d+[^.]*\.', PROMPT)
print(json.dumps({"clauses": clauses[:10], "total": len(clauses)}))

bibividi add  scripts/my-skill/    # install
bibividi list                       # inspect
bibividi test my-skill-id           # run against stdin

Sandbox: only json, re, math, statistics, datetime, collections, itertools, functools, pathlib, string, textwrap, csv, io, base64, hashlib, uuid are allowed. No OS, network, or subprocess access. Scripts are SHA-256 verified on install. Ed25519 signatures supported for marketplace packs.

Included skill library (26 packs)

Demo packs

Pack	Typical saving
pdf-clause-extractor	80–95%
csv-data-summarizer	85–95%
structured-report-extractor	75–90%
batch-translation-condenser	70–90%
code-review-condenser	60–80%

Office productivity

Pack	What it extracts	Typical saving
email-thread-condenser	Action items, decisions, timeline from email threads	75–90%
meeting-notes-structurer	Agenda, decisions, action items from meeting notes	70–85%
slack-thread-distiller	Key decisions and outcomes from Slack threads	70–88%
cv-profile-extractor	Skills, experience, education from CVs/résumés	75–90%
job-posting-parser	Role, requirements, salary from job descriptions	70–85%
expense-report-analyzer	Totals, categories, anomalies from expense reports	80–90%
invoice-parser	Line items, totals, parties from invoices	80–92%
sales-pipeline-condenser	Stage breakdown, top deals, at-risk from CRM exports	75–90%
customer-feedback-synthesizer	Sentiment, themes, praise/complaints from reviews	80–92%
requirements-to-stories	User stories grouped by epic from PRDs/specs	60–80%
project-status-condenser	RAG status, completed/blocked/next-steps from status reports	70–85%
legal-contract-summarizer	Parties, term, obligations, key clauses from contracts	85–95%
support-ticket-analyzer	Priority distribution, categories, critical tickets from ticket batches	75–90%

BSS 1.1 chain pipelines

Pack	Pipeline	Typical saving
document-intelligence-suite	pdf-text-extractor → type-classifier → content-router	90–97%
financial-pipeline	invoice-parser → line-validator → totals-summarizer	95–99%
code-review-pipeline	diff-parser → complexity-scorer → review-formatter	92–98%

---

Multi-provider support

BIBIVIDI works with any OpenAI-compatible API out of the box. Supported by default:

Provider	Host	Notes
Anthropic	api.anthropic.com	Full support
OpenAI	api.openai.com	Full support
Google Gemini	generativelanguage.googleapis.com	Full support incl. contents[] format
Groq	api.groq.com	Free tier — quota savings mode
Mistral	api.mistral.ai	OpenAI-compatible
Together.ai	api.together.xyz	OpenAI-compatible
Ollama	localhost:11434	Requires BIBIVIDI_ALLOW_LOCAL_APIS=true

Free-tier / quota mode — when model_price_eur = 0 (Gemini free, Groq, Ollama), BIBIVIDI switches framing from "€ saved" to "extra requests/day":

bibividi status
  Quota saved : +850 extra prompts/day within your Groq free tier

---

BSS 1.1 — Skill chains

Chain multiple skills into a typed I/O pipeline. Each skill's output becomes the next skill's input:

bibividi chain build   # scaffold a chain pack from installed skills
bibividi chain ...     # manage chains

Enable in config:

BIBIVIDI_ENABLE_CHAINS=true
BIBIVIDI_MAX_CHAIN_DEPTH=10        # max skills in a chain
BIBIVIDI_MAX_CHAIN_RUNTIME_S=0     # 0 = 3× MAX_SCRIPT_RUNTIME_S

BSS 2.0 — Auto-assembly

Automatically discovers and assembles typed I/O pipelines from installed skills without manual chain definitions:

BIBIVIDI_ENABLE_AUTO_ASSEMBLY=true

---

HTTPS interception (optional)

For apps that send directly to https://api.anthropic.com without going through the SDK integration above:

bibividi cert generate   # create local CA + server cert
bibividi cert install    # trust the CA (may require sudo)
bibividi start --ssl     # HTTP :8080 + HTTPS CONNECT tunnel :8443

Configure your HTTP client to use http://127.0.0.1:8443 as a proxy:

import httpx
client = httpx.Client(proxy="http://127.0.0.1:8443", verify="~/.bibividi/ca.crt")

---

Auto-generate skills (generator)

When no script matches a prompt, BIBIVIDI calls the LLM once to generate a new BSS skill automatically. The next similar request will be handled locally. Generated skills are held for review until approved.

BIBIVIDI_GENERATOR_ENABLED=true          # on by default
BIBIVIDI_GENERATOR_API_KEY=sk-ant-...    # key used only for generation
BIBIVIDI_GENERATOR_MODEL=claude-haiku-4-5-20251001
BIBIVIDI_GENERATOR_PROVIDER=anthropic    # anthropic|openai

Review and approve generated skills:

bibividi review          # list pending skills
bibividi approve <id>    # activate a skill for execution

---

CLI reference

Interactive shell (REPL)

Running bibividi with no arguments enters an interactive shell — no need to prefix every command:

$ bibividi
BIBIVIDI — AI token optimization proxy
Type help or ? for commands  ·  exit or Ctrl+D to quit

bibividi> start
Proxy started on 127.0.0.1:8080
Point your apps to http://127.0.0.1:8080  ·  'tail' to watch requests  ·  'stop' to stop

bibividi> status
BIBIVIDI Status
  Library : 3 scripts
  ...

bibividi> list
bibividi> tail
bibividi> stop
bibividi> exit
Bye!

Shell command	Description
start [--ssl]	Start the proxy in the background (shell stays live)
stop	Stop the running proxy
status	Show proxy status and token savings
list [--chains]	List installed skills
add <pack-dir>	Install a BSS skill pack
review	List pending auto-generated skills
approve <id>	Activate a pending skill
tail	Stream live request events (Ctrl+C to return to shell)
simulate <file>	Dry-run against a request log
calibrate [file]	Tune the match threshold
demo	Run the built-in token savings demo
digest [--now]	Show weekly savings digest
report	CSRD AI Scope 3 emissions report
roi	Open ROI calculator in browser
feedback bad <id>	Flag a bad skill output
activate <key>	Activate a license key
help / ?	List all commands
exit / Ctrl+D	Quit the shell

Command history is persisted across sessions in ~/.bibividi/shell_history.

One-liner mode

All commands also work directly without entering the shell:

# Core
bibividi activate <license-key>     # activate your license (one-time)
bibividi start [--ssl] [--host HOST] [--port PORT]
bibividi status
bibividi list
bibividi add <pack-dir>
bibividi test <script-id>           # prompt via stdin
bibividi demo                       # in-process demo on a sample contract

# Certificates
bibividi cert generate|install|path

# Skill chains (BSS 1.1)
bibividi chain build

# Approval workflow (auto-generated skills)
bibividi review
bibividi approve <skill-id>

# Trust feedback
bibividi feedback ...               # flag bad skill outputs

# Observability
bibividi tail                       # stream live request events
bibividi calibrate [prompts-file]   # tune match threshold
bibividi simulate <requests-file>   # dry-run against a request log

# Reporting & billing
bibividi digest                     # weekly savings digest
bibividi digest-quarterly           # send quarterly email digest
bibividi report                     # CSRD AI Scope 3 emissions report
bibividi roi                        # open ROI calculator in browser
bibividi billing ...
bibividi marketplace ...

Configuration

All settings via BIBIVIDI_ environment variables (or .env):

Variable	Default	Description
BIBIVIDI_PROXY_PORT	8080	Proxy listen port
BIBIVIDI_TUNNEL_PORT	8443	HTTPS CONNECT tunnel port
BIBIVIDI_MATCH_THRESHOLD	0.75	Cosine similarity threshold
BIBIVIDI_MAX_SCRIPT_RUNTIME_S	10.0	Sandbox timeout
BIBIVIDI_DRY_RUN	false	Log savings without injecting condensed result
BIBIVIDI_SLACK_WEBHOOK	—	Post savings alerts to Slack
BIBIVIDI_GENERATOR_ENABLED	true	Auto-generate skills on no-match
BIBIVIDI_GENERATOR_MODEL	claude-haiku-4-5-20251001	Model used for generation
BIBIVIDI_GENERATOR_PROVIDER	anthropic	anthropic or openai
BIBIVIDI_ENABLE_CHAINS	false	Enable BSS 1.1 skill chains
BIBIVIDI_MAX_CHAIN_DEPTH	10	Max skills per chain
BIBIVIDI_ENABLE_AUTO_ASSEMBLY	false	Enable BSS 2.0 auto-assembly
BIBIVIDI_ALLOW_LOCAL_APIS	false	Allow localhost targets (Ollama)
BIBIVIDI_QUOTA_DAILY_TOKENS	—	Override daily token quota for quota-mode display
BIBIVIDI_BILLING_TIER	pro	free | pro | growth | enterprise
BIBIVIDI_BILLING_API_URL	—	URL of the billing/license API
BIBIVIDI_LOG_REQUESTS	false	Log prompt content (dev only)

---

Development

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env

pytest                    # 294 tests
ruff check .
mypy bvd/

# Zero-credit local testing
uvicorn mock_api:app --port 9090 &
BIBIVIDI_TARGET_APIS='["localhost:9090"]' bibividi start

---

Releasing

git tag v1.2.3
git push origin v1.2.3

This triggers two workflows automatically:

• Release — builds Tauri desktop installers (Linux .AppImage/.deb, macOS .dmg, Windows .msi) and creates a draft GitHub release
• Publish to PyPI — builds and publishes the Python package via OIDC trusted publishing

To publish the GitHub release, go to Releases, review the draft, and click Publish release.

---

License

MIT — engine is open source. Cloud runner and marketplace are proprietary (Pro/Enterprise).