personakit 0.2.1

A declarative framework for building role-based LLM agents — code reviewers, compliance officers, clinical triage, support triage. Not a personality classifier; an agent builder. Persona, frameworks, probes, red flags, and themes as data; you get a typed, cited, safety-aware agent.

A declarative framework for building role-based LLM agents.
Describe a specialist as data — persona, frameworks, probes, red flags, themes —
and get a typed, cited, safety-aware agent. No chain wiring. No graph building.

Why · Quickstart · Specialists · Providers · FAQ · GitHub

pip install personakit

Created by Majidul Islam · MIT licensed · Independent open source

---

What personakit is — and what it is not

personakit IS a declarative framework for building role-based LLM agents: code reviewers, compliance officers, clinical triage, support triage, contract reviewers, scrum masters, and similar domain specialists. You describe the role as data; personakit produces the runnable agent backed by OpenAI / Anthropic / your local model.

Production wiring shipped in v0.2:

• ✅ Native streaming with Agent.analyze_stream() (OpenAI / Anthropic / LiteLLM / Mock)
• ✅ OpenTelemetry hooks via the Tracer protocol — analyze / provider / tool spans
• ✅ Token cost tracking — result.estimated_cost_usd with pricing for ~25 popular models
• ✅ Multi-turn conversations with ConversationalAgent + serialisable Session
• ✅ Multi-turn tool-calling loop — same Specialist works on every provider
• ✅ Real-time web knowledge via personakit[web] (fetch_url, tavily_search, …)

personakit is NOT:

• ❌ Not a personality classifier (not MBTI, not Big Five, not trait inference). It has nothing to do with pypersonality, persai, or similar trained classifiers.
• ❌ Not an ML training library. No feature extraction, no fitted models, no datasets. It wraps LLMs you already have access to.
• ❌ Not a RAG framework. Bring your own vector store via the optional @tool system — we don't ship embeddings or retrieval.
• ❌ Not a chain-orchestration engine. Composable alongside LangChain, LangGraph, and CrewAI — personakit owns the specialist layer, they own the pipeline layer.

---

Why personakit?

Building a specialist LLM agent shouldn't take 200 lines of chain wiring.

Every specialist — regardless of domain — has the same anatomy:

• A role the agent plays
• Knowledge bodies it draws from and cites
• Diagnostic questions it asks of any input
• Safety triggers demanding immediate action
• Output categories organizing what it recommends

That anatomy holds whether your specialist reviews legal documents, evaluates clinical cases, audits financial transactions, scores research papers, drafts user stories, scopes engineering work, supports customers, moderates content, qualifies sales leads, grades coursework, or any of a thousand other roles. personakit captures the anatomy. You bring the role.

You describe WHO the specialist is, as data. The library produces a typed, cited, safety-aware agent that runs on any LLM provider — without chain wiring, graph building, or orchestration code.

The discipline is the role description. Everything else — JSON schema generation, red-flag matching (deterministic + semantic), citation enforcement, provider routing, structured-output validation — is automatic.

If a domain expert can articulate what they look for, what they ask, and what they recommend, you can build them an agent in 30 lines.

The primitives

Concept	What it gives you
Specialist	Frozen dataclass — the entire agent definition, authorable in YAML
Framework	Body of knowledge with a citation key, cited in output
Probe	Diagnostic question; becomes a typed field in the structured response
RedFlag	Trigger → severity → action → citation, matched deterministically AND semantically
Theme	User-selectable recommendation category
Priority	Always-on checks reported as met / unmet / unknown
Tool (optional)	@tool decorator — opt-in for external memory, DB, APIs

Core has just two runtime deps: pydantic and httpx.

---

30-second quickstart

import asyncio
from personakit import Agent, Specialist, Framework, Probe, RedFlag, Severity, Theme

code_reviewer = Specialist(
    name="code_reviewer",
    persona="Senior staff engineer. Correctness, security, perf — in that order.",
    frameworks=[Framework(name="OWASP Top 10"), Framework(name="SOLID")],
    probes=[Probe(question="Does the change include tests?", value_type="boolean")],
    red_flags=[
        RedFlag(
            trigger="Hard-coded secret",
            severity=Severity.CRITICAL,
            action="BLOCK merge. Rotate the secret immediately.",
            patterns=[r"sk-[A-Za-z0-9]{20,}", r"AKIA[0-9A-Z]{16}"],
        ),
    ],
    themes=[Theme(name="Correctness"), Theme(name="Security"), Theme(name="Performance")],
)

agent = Agent(specialist=code_reviewer, model="gpt-4o-mini")
result = asyncio.run(agent.analyze("--- a.py\n+api_key = 'sk-proj-abc123456789012345678'"))

print(result.pretty())              # full structured summary
print(result.red_flags_triggered)   # [TriggeredRedFlag(..., severity=CRITICAL, evidence='sk-proj-...')]
print(result.has_urgent)            # True

That's the whole agent. No chain wiring. One Specialist dataclass. Typed, cited, safety-aware output.

---

Production-ready in v0.2

Four features that turn personakit from "interesting alpha" into something you can deploy:

Streaming — agent.analyze_stream(text)

async for event in agent.analyze_stream(case_text):
    if event.type == "text_delta":
        print(event.text, end="", flush=True)
    elif event.type == "red_flag_pre_match":
        print(f"\n🚨 {event.red_flag.trigger}")
    elif event.type == "complete":
        result = event.result   # full AnalyzeResult

Live event stream with deterministic red flags fired up front, text deltas as the LLM types, tool-call lifecycles, and a final structured result. Native streaming on OpenAI, Anthropic, LiteLLM (100+ providers), and MockProvider.

OpenTelemetry hooks — personakit.observability

from personakit.observability import OpenTelemetryTracer
agent = Agent(specialist=..., model="gpt-4o", tracer=OpenTelemetryTracer())

Three span points: personakit.analyze, personakit.provider.complete, personakit.tool.invoke. Plug in your existing OTel pipeline (LangSmith, Datadog, Honeycomb, Jaeger) — or implement the Tracer Protocol in 30 lines for any other backend. Install: pip install 'personakit[otel]'.

Token cost tracking — result.estimated_cost_usd

result = await agent.analyze(text)
print(result.usage)              # {"prompt_tokens": 1200, "completion_tokens": 350}
print(result.estimated_cost_usd) # 0.00465

Pricing tables for ~25 models (OpenAI, Anthropic, Gemini, Groq, DeepSeek, Mistral). Local models (Ollama, vLLM) cost 0.0. Unknown models return None so callers can distinguish "unknown" from "free". Add custom rates with register_pricing(...).

Conversational sessions — ConversationalAgent + Session

from personakit import ConversationalAgent
from personakit.examples import CUSTOMER_SUPPORT_TRIAGE

agent = ConversationalAgent(specialist=CUSTOMER_SUPPORT_TRIAGE, model="gpt-4o-mini")
session = agent.start_session(user_id="alice")

reply1 = await session.send("My order ORD-1002 is late")
reply2 = await session.send("It's been 3 weeks now")  # remembers turn 1

# Caller-managed persistence — serialise to your choice of store
blob = session.serialize()
restored = Session.deserialize(blob, agent=agent)

Multi-turn memory with a configurable history window. No database required — sessions serialise to a string the caller can stick in Redis, Postgres, or a JSON file.

---

Bundled specialists — 7 domains, zero boilerplate

Import any of these directly, or read the source as a template:

Specialist	Domain	What it does
CODE_REVIEWER	engineering.software.review	PR reviewer — OWASP, SOLID, 12-Factor, secret detection
FINTECH_TRANSACTION_REVIEWER	finance.fintech.aml	AML/fraud triage — OFAC, FATF typologies, SAR filing
CUSTOMER_SUPPORT_TRIAGE	support.saas.b2c	SaaS B2C support — refund policy, chargeback, GDPR routing
SCRUM_MASTER	engineering.delivery.agile	Sprint health — scope creep, WIP limits, blockers
CONTRACT_REVIEWER	legal.contracts.m_and_a	M&A redlining — English common law, UCC, GDPR Art. 28
FALLS_PREVENTION_NURSE	healthcare.clinical.falls	Post-fall clinical triage — NICE NG161, CG176, Morse
MATH_TUTOR	education.secondary	Socratic GCSE tutor — minimal persona-only specialist

from personakit import Agent
from personakit.examples import FINTECH_TRANSACTION_REVIEWER

agent = Agent(specialist=FINTECH_TRANSACTION_REVIEWER, model="gpt-4o-mini")
result = asyncio.run(agent.analyze(transaction_details))

---

Real agent types you can build in one file

You want an agent that...	Define these concepts
Reviews pull requests for security and correctness	Framework(OWASP), RedFlag(sql_injection, hardcoded_secret), Theme(Security, Performance)
Screens fintech transactions for AML / sanctions	Framework(BSA/AML, OFAC), RedFlag(sanctioned_counterparty, structuring), typology Themes
Triages customer support messages	Probe(order_id, sentiment), RedFlag(chargeback_language), Theme(Resolution, Escalation)
Coaches a sprint team as a scrum master	Framework(Scrum Guide, DORA), RedFlag(scope_creep, external_blocker), Theme(At-risk stories, Retro candidates)
Reviews M&A contracts for legal risk	Framework(English common law, UCC), RedFlag(unlimited_liability), Theme(Liability, IP)
Runs a post-fall clinical assessment	Framework(NICE NG161, NICE CG176), RedFlag(LOC, head_contact_on_anticoagulant), clinical probes
Writes product specs against JTBD	Framework(Jobs-to-be-Done, RICE), RedFlag(no_success_metric), Theme(Edge cases, Open questions)
Tutors a student without giving away answers	Theme(Concept check, Next hint, Common pitfall), Constraint(no direct answer first)
Does equity research on a public SaaS name	Framework(DCF, Rule of 40), RedFlag(NDR_below_100, negative_fcf_plus_decel), Theme(Thesis, Risks)

YAML authoring — hand off to a domain expert

name: code_reviewer
persona: Senior staff engineer reviewing PRs. Correctness, security, perf, in that order.
frameworks: [OWASP Top 10, SOLID, 12-Factor App]
probes:
  - question: Does the change include tests?
    key: has_tests
    value_type: boolean
    weight: high
red_flags:
  - trigger: Hard-coded secret or API key
    severity: critical
    action: BLOCK merge. Rotate the secret. Move to a secret manager.
    match: both
    patterns: ['sk-[A-Za-z0-9]{20,}', 'AKIA[0-9A-Z]{16}']
themes: [Correctness, Security, Performance, Maintainability, Tests]

from personakit import Specialist, Agent

spec = Specialist.from_yaml("code_reviewer.yaml")
agent = Agent(specialist=spec, model="claude-sonnet-4-6")

Red flags — the feature no-one else has

Every RedFlag is a trigger → severity → action → citation contract, matched in two phases:

1. Deterministic pre-match — regex / keywords on raw input. Fast, offline, quotable.
2. Semantic post-match — the LLM evaluates whether the trigger applies in context. Catches paraphrase, synonyms, implicit meaning.

Results merge, with deterministic evidence winning on ties:

RedFlag(
    trigger="Hard-coded secret, token, or credential",
    severity=Severity.CRITICAL,
    action="BLOCK merge. Rotate secret. Move to secret manager.",
    citation="OWASP A02:2021",
    match=MatchMode.BOTH,
    patterns=[r"sk-[A-Za-z0-9]{20,}", r"AKIA[0-9A-Z]{16}"],
)

Structured output — derived from the Specialist

You never write a JSON schema by hand. The probes, red flags, and themes are the schema:

result = await agent.analyze(input_text)

result.summary                # narrative summary
result.probes_answered        # {probe_key: typed_value_or_null}
result.probes_unanswered      # list[Probe] — next-round questions
result.red_flags_triggered    # list[TriggeredRedFlag] with evidence + source
result.recommendations        # themed list with citations
result.citations_used         # framework citation keys referenced
result.priorities_status      # per-priority met / unmet / unknown
result.has_urgent             # convenience boolean

Tools — opt-in, with a real multi-turn loop

Core has zero coupling to tool calling. When you want tools, decorate functions and attach them to an Agent:

from personakit.tools import tool

@tool
def lookup_order(order_id: str) -> dict:
    """Fetch an order from the order database."""
    return order_db.get(order_id)

@tool
def search_knowledge_base(query: str, top_k: int = 5) -> list[str]:
    """Semantic search against the internal KB — your vector store."""
    return kb.search(query, k=top_k)

agent = Agent(specialist=CUSTOMER_SUPPORT_TRIAGE, model="gpt-4o-mini")
agent = agent.with_tools([lookup_order, search_knowledge_base])

result = await agent.analyze("Where is order ORD-1002?")
# Behind the scenes:
#   1. LLM emits tool_calls in its response
#   2. personakit invokes the matching tool locally
#   3. The result is fed back into the conversation
#   4. The LLM produces the final structured analysis

The loop runs across OpenAI, Anthropic, LiteLLM, and MockProvider identically — personakit normalises between OpenAI's tool_calls array and Anthropic's tool_use / tool_result content blocks internally. Schemas are auto-built from your function signatures and docstrings.

Bound the loop with Agent(..., max_tool_iterations=6) to cap cost; defaults to 6.

Real-time knowledge from URLs (personakit[web])

For the common case "I want my agent to use a web link as its knowledge source", personakit ships a small set of ready-made tools:

pip install 'personakit[web]'

from personakit import Agent
from personakit.web import fetch_url, extract_article, tavily_search
from personakit.examples import FINTECH_TRANSACTION_REVIEWER

Pattern A — pre-fetch (deterministic, single LLM call)

Fetch the URL yourself, pass the content as extra_context. Best when you know the URL up front (e.g. a user submits a link).

fetched = await fetch_url.invoke(url="https://www.reuters.com/some-article")
result = await agent.analyze(
    "Is the entity in this press release on any sanctions list?",
    extra_context=f"Source: {fetched['final_url']}\n\n{fetched['text']}",
)

Pattern B — LLM decides when to fetch (autonomous)

Attach the tools and let the agent decide. Best when the agent might need multiple sources or doesn't know in advance whether to fetch.

agent = (
    Agent(specialist=FINTECH_TRANSACTION_REVIEWER, model="gpt-4o-mini")
    .with_tools([fetch_url, tavily_search])
)

result = await agent.analyze(
    "Verify this counterparty against current sanctions lists: ACME Holdings Pte Ltd"
)
# The LLM may call tavily_search first, then fetch_url on a result link,
# then return the final analysis. The tool loop runs all of it.

Available web tools

Tool	Purpose	Requirements
fetch_url(url, max_chars=8000)	HTTP GET + text extraction (BeautifulSoup)	personakit[web]
extract_article(url, max_chars=12000)	Smarter article extraction (trafilatura)	personakit[web]
tavily_search(query, max_results=5)	LLM-optimised web search	TAVILY_API_KEY env var
serper_search(query, max_results=5)	Google SERP search	SERPER_API_KEY env var

Both Tavily and Serper offer free tiers (1,000 / 2,500 searches/month). All four work cross-provider — same agent, same tool list, any LLM backend.

Registry — one app, many specialists

from personakit import SpecialistRegistry

registry = SpecialistRegistry.from_directory("personas/")

# Route by domain
engineering_agents = registry.by_domain("engineering")
finance_agents = registry.by_domain("finance")

# Route by name
support = registry.get("support_triage")

Providers

Extra	Install	What you get
personakit[openai]	openai>=1.0	Native OpenAI SDK — gpt-4o-mini default
personakit[anthropic]	anthropic>=0.20	Native Anthropic SDK — claude-sonnet-4-6
personakit[litellm]	litellm>=1.40	100+ providers via one extra (see below)
personakit[web]	beautifulsoup4>=4.12, trafilatura>=1.6	URL fetch + article extraction + search
personakit[otel]	opentelemetry-api>=1.20, -sdk>=1.20	OpenTelemetry tracer adapter
personakit[yaml]	pyyaml>=6.0	YAML specialist authoring
personakit[all]	all of the above	Everything

100+ providers via LiteLLM

LiteLLM normalises the APIs of 100+ providers — OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex AI, Cohere, Mistral, Hugging Face, Ollama, DeepSeek, Together AI, Groq, Fireworks, Anyscale, and any OpenAI-compatible endpoint — into a single unified call. LiteLLMProvider plugs that into personakit so switching providers is a one-line change:

from personakit import Agent
from personakit.providers import LiteLLMProvider
from personakit.examples import FINTECH_TRANSACTION_REVIEWER

# Same Specialist, any provider LiteLLM supports — change the model string only.
provider = LiteLLMProvider(default_model="bedrock/anthropic.claude-v2")
# or:  LiteLLMProvider(default_model="azure/my-gpt-4-deployment",
#                       api_key=..., api_version="2024-06-01")
# or:  LiteLLMProvider(default_model="vertex_ai/gemini-pro")
# or:  LiteLLMProvider(default_model="ollama/llama3",
#                       api_base="http://localhost:11434")
# or:  LiteLLMProvider(default_model="groq/mixtral-8x7b-32768")

agent = Agent(specialist=FINTECH_TRANSACTION_REVIEWER, provider=provider)
result = await agent.analyze(transaction_details)

Install: pip install 'personakit[litellm]'. You can also use LiteLLM's proxy mode with OpenAIProvider(base_url="http://localhost:4000") if you prefer running LiteLLM as a separate gateway.

MockProvider ships in the core for offline testing — no API key needed:

from personakit.testing import MockProvider
provider = MockProvider(responses={"summary": "...", "recommendations": [...]})

Testing helpers

from personakit.testing import assert_triggered, assert_cited

result = await agent.analyze("Customer: 'I want to chargeback this transaction.'")
assert_triggered(result, "legal_or_chargeback_language_attorney_lawsuit_chargeback_small_claims_bbb")

Design principles

1. Specialist is pure data. No behaviour, no side effects, serializable.
2. Schema is derived. Probes, red flags, themes are the output contract.
3. Deterministic where possible, semantic where needed. Red flags run both.
4. Tools are opt-in. Core has zero coupling to tool calling.
5. Minimal dependencies. pydantic + httpx. Everything else is an extra.
6. Domain-neutral. Engineering, support, fintech, legal, clinical, education, delivery, product — one library.
7. Provider-agnostic. Native OpenAI + Anthropic adapters, plus 100+ providers via the LiteLLM extra (Azure, Bedrock, Vertex AI, Cohere, Mistral, Ollama, Groq, and any OpenAI-compatible endpoint). Same Specialist, any model.

Works alongside the rest of the LLM toolchain

personakit focuses on declarative specialist definition. It intentionally does not try to be:

• a chain-composition library — use LangChain when you need to wire up complex multi-step LLM pipelines
• a multi-agent orchestration framework — use CrewAI when you need a team of agents collaborating on a shared goal
• a branching control-flow engine — use LangGraph when you need conditional routing and loops across nodes

They compose nicely. A LangChain chain can invoke a personakit Agent as one of its steps. A CrewAI crew member can be a personakit Specialist. A LangGraph node can call agent.analyze() and route on result.has_urgent.

Use personakit for what it's best at — the declarative specialist layer — and reach for the others when the problem actually needs chains, crews, or graphs.

---

FAQ

Q: Is this a personality classifier (MBTI, Big Five, etc.)? No. personakit is an agent builder. It has no trained models, no feature extraction, and no personality taxonomy. If you need MBTI or Big Five, look at pypersonality or persai — completely different category of library.

Q: Can it fetch external knowledge (RAG, vector store, real-time APIs)? Yes — via the opt-in @tool system. personakit is bring-your-own-retrieval: wrap your vector store (Pinecone, pgvector, Chroma, Qdrant, Weaviate) or any API in a @tool function, and the agent uses it. We don't ship a vector store because every team already has one they prefer.

Q: Why not just use LangChain or LangGraph? Different problems. LangChain/LangGraph describe what the agent does (imperative chains, graphs). personakit describes who the agent is (declarative role). For role-based agents the declarative approach is ~10× less code and lets non-engineers author specialists in YAML. Use personakit inside a LangChain chain or LangGraph node when you need the chain / graph for orchestration.

Q: Does it work with LiteLLM (Azure, Bedrock, Vertex AI, Ollama, Groq, …)? Yes — install with pip install 'personakit[litellm]' and use LiteLLMProvider(default_model="bedrock/anthropic.claude-v2") (or any LiteLLM model string). The LiteLLMProvider adapter exposes 100+ providers through the same Agent interface as the native OpenAI / Anthropic adapters.

Q: What does it depend on? Just two runtime dependencies: pydantic and httpx. Providers (openai, anthropic, litellm, pyyaml) ship as optional extras — install only what you need. Total transitive footprint on a fresh venv with no extras: ~12 packages.

Q: Do I need to be a Python expert to use it? No. The minimum useful Specialist is a name + a persona string. You can declare more capable specialists in pure YAML — no Python required for the authoring side. An engineer plugs the YAML into the runtime in 2 lines of code.

Q: Is it production-ready? It's v0.2.0 — alpha, but with the production basics in place: streaming (Agent.analyze_stream()), OpenTelemetry hooks (personakit[otel]), token cost tracking (result.estimated_cost_usd), and conversational sessions (ConversationalAgent). 93 tests pass; mypy --strict clean across 29 source files; no unreleased breaking changes. The API may still evolve before v1.0 — pin the minor version in production until then.

Q: Does it support streaming? Yes. Agent.analyze_stream(text) returns an async iterator yielding StreamEvent objects: red_flag_pre_match, text_delta, tool_call, tool_result, iteration_complete, complete, error. Streaming works natively on OpenAI, Anthropic (with tool_use / tool_result content-block translation), LiteLLM (100+ providers), and MockProvider.

Q: How do I integrate with my existing observability stack (LangSmith / Datadog / Honeycomb / Jaeger)? Plug in OpenTelemetryTracer (install via personakit[otel]) — three spans (personakit.analyze, personakit.provider.complete, personakit.tool.invoke) flow into your existing OTel pipeline. Or implement the Tracer Protocol in ~30 lines for any other backend.

Q: Does it track token cost? Yes. AnalyzeResult.estimated_cost_usd returns a USD float for the ~25 models in the bundled pricing table (OpenAI, Anthropic, Gemini, Groq, DeepSeek, Mistral). Local models (Ollama / vLLM) cost 0.0. Unknown models return None so callers can distinguish "unknown" from "free". Add custom rates with register_pricing(...).

Q: How do I do multi-turn conversation? Use ConversationalAgent + Session. The session tracks history; each session.send(message) includes the prior turns as context. Persist sessions yourself (session.serialize() returns a JSON blob) — no DB requirement.

Q: Is there commercial support / a hosted offering? No. personakit is an independent open-source project — MIT licensed, no SaaS tier, no telemetry, no upsell path. Use it freely.

Q: Can I contribute? Yes — see Contributing below. Bug reports, feature requests, new bundled specialists, and PRs all welcome.

---

Contributing

personakit is a solo independent project — every contribution counts.

Quick start

git clone https://github.com/Majidul17068/personakit.git
cd personakit
python -m venv .venv && source .venv/bin/activate
pip install -e '.[dev,openai,anthropic,litellm,yaml,web,otel]'

Quality gates (all must pass before opening a PR)

pytest                # unit tests — currently 93 passing
mypy --strict src     # zero errors across 29 source files
ruff check src tests  # lint
python -m build       # wheel + sdist build cleanly

What's most useful

• New bundled specialists for domains we don't yet cover (src/personakit/examples/). See code_reviewer.py or fintech_reviewer.py as templates.
• Bug reports with a minimal reproduction — open an issue.
• Real-world API feedback — what's clunky, what's missing, where it breaks.
• Documentation improvements — clarifications, fixes, examples.

PRs land faster when they include a test for the change.

---

Privacy

personakit does not collect telemetry. No network calls outside the LLM provider you configure. No analytics, no anonymous usage statistics, no phoning home. The library is a thin layer over your own LLM API key.

---

Status

v0.2.0 — alpha. All four production basics shipped (streaming, OpenTelemetry, cost tracking, conversational sessions). 93 tests passing, mypy --strict clean across 29 source files. API may still evolve before v1.0 — pin the minor version in production. See CHANGELOG.md for the full release history and ROADMAP.md for what's next.

Author

Majidul Islam — @Majidul17068

personakit is an independent open-source project. Contributions welcome.

License

MIT © 2026 Majidul Islam.