TARA — Tax, Assets & Residency Advisor

TARA answers a focused question: a rule changed—who is affected, what should they do, when, and what evidence would close the work?

flowchart LR
    A[Rule changes] --> B[LLM understands the request]
    B --> C[MCP discovers TARA tools]
    C --> D[Deterministic TARA checks source, dates and facts]
    D --> E{Information missing?}
    E -->|Yes| F[Ask; do not guess]
    E -->|No| G[Create a dated action plan]
    G --> H{Is the evidence correct?}
    H -->|No| I[Return it with a clear reason]
    H -->|Yes| J[Ready for human review]
    C & D & G & H --> K[(Keep an inspectable proof record)]

Start here

1. Revenue guidance changes the fund-tax rate in Section 4.3 from 41% to 38% for deemed disposals arising on or after 1 January 2026.

2. TARA shows the old wording and the new wording side by side, while keeping digital fingerprints of both sources.

3. It preserves the former and current rule instead of overwriting history.

4. It uses the holding's event date to select 41% before 2026 or 38% from 2026.

5. It creates a clear action plan with dates and the evidence to keep.

6. It rejects a calculation using the old 41% rate and accepts an exact 38% calculation.

7. It verifies the linked proof record for the run.

The browser labels Maeve as the recommended path. Ciarán, Priya, and Arun are exploratory breadth cases and must not be described as independently validated legal advice.

Related MCP server: AEGIS Governance

Why this is an AI agent, not a chatbot

• MCP makes the AI inspectable. The browser exposes a readable trace of the model's tool calls rather than asking a judge to trust a fluent answer.

• Models never own decisive calculations. Dates, diffs, effective-date selection, threshold checks, arithmetic, action ordering, and evidence outcomes are deterministic Python.

• Rules live in data. Domain packs define sources, obligations, scoping, triggers, obligation-level predicates, and evidence rules.

• Missing material facts fail closed. An incomplete threshold or event becomes indeterminate; COURSE creates no action.

• Pack scope is not obligation scope. COMPASS can confirm that a rule family is relevant while PLOT rules individual duties in, out, or indeterminate.

• History is inspectable. Every agent writes to ATLAS, an append-only JSONL chain with linked hashes.

Current evidence

At the latest verification:

• 115 automated tests pass.

• 8 executable acceptance checks pass in tools/build_evaluation_card.py.

• 9 domain packs are loaded: six standalone packs and three declarative cross-border corridors.

• 12 MCP tools expose the system, including discovery for domains, sources, and prepared holdings.

• The browser demo defaults to a server-side OpenAI orchestrator and an inspectable MCP trace; deterministic calculation and labelled replay provide fallback layers.

• Replay output is generated from real API responses for network-safe presentation fallback.

Open band — what the rule says

Tenant band — what the rule means for one case

Domain coverage

The project loads six standalone packs:

1. Irish fund taxation and eight-year deemed disposal (tax)

2. Indian foreign-asset and foreign-income disclosure (india-fa)

3. US foreign-account and foreign-asset reporting (us-fbar)

4. Irish residence-permit registration and renewal (ireland-irp)

5. Irish capital-gains tax on property (ireland-cgt-property)

6. Indian NRI listed-securities capital gains (india-nri-securities)

It also synthesizes three corridor packs from domains/interactions.yaml:

• India–Ireland

• India–United States

• Ireland–United States

Coverage is selected prototype coverage, not comprehensive compliance clearance.

Obligation-level safety

A pack-level determination is intentionally coarse. For example, a US tax resident may place the US pack in scope, but that alone must not create FBAR, Form 8938, Form 3520, and Form 8621 actions for every asset.

Each ObligationSpec can therefore declare:

• effective_from and effective_to

• applies_when predicates using deterministic operators such as equals, in, gt, and lte

PLOT reports one of:

Repository map

domains/                         cited rules, scope, triggers, and evidence contracts
data/sources/                     controlled v1/v2 source snapshots
registers/holder_register.json    synthetic test register
tara/core/                        dates, diffs, pack and corridor loaders
tara/agents/                      the nine agents
tara/atlas/                       append-only hash-linked ledger
tara/mcp_server/                  twelve MCP tools over stdio or streamable HTTP
tara/orchestrator/                deterministic and optional LLM clients
tara/pipeline.py                  band-order-enforcing orchestration
demo/                             FastAPI adapter, browser UI, presets, replay
tests/                            unit, integration, safety, and end-to-end tests
tools/build_evaluation_card.py    executable public acceptance report
docs/tara-project-guide.html      canonical visual and technical walkthrough
public-release-manifest.txt       allowlist for clean public publication

Run locally

python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
pip install -r demo/requirements.txt

pytest -q
python tools/build_evaluation_card.py
python -m uvicorn demo.api:app --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000, choose Maeve, and follow the source change to the evidence check.

MCP interface

python -m tara.cli serve

TARA exposes twelve tools:

• list_domains

• list_sources

• list_holdings

• survey_detect_change

• legend_decompose

• almanac_refresh

• compass_assess

• plot_align

• course_plan

• anchor_verify

• meridian_survey

• atlas_reconstruct

A streamable-HTTP server is also available:

python -m tara.cli serve --transport streamable-http --port 8000

With the llm extra installed and OPENAI_API_KEY configured, run the real OpenAI-over-MCP path:

pip install -e '.[llm]'
python -m tara.cli orchestrate \
  --model gpt-4.1-mini --show-tool-calls

The model discovers exact domain, source, and holding IDs through MCP tools. If a required applicability fact is missing, it stops and asks for that fact rather than guessing. The browser runs the same LLM-first sequence server-side, then presents the independently derived deterministic controls and a readable MCP trace.

Demo and replay

Regenerate replay data after changing packs, register fields, agent behavior, or presets:

python -m uvicorn demo.api:app --port 8000 &
python demo/capture_replay.py

Public release without private history

The private repository remains the working source of truth. The public version is built from an explicit allowlist into a separate directory:

python tools/build_public_release.py

1. Copies only paths in public-release-manifest.txt.

2. Omits .git, internal reviews, generated state, caches, local environments, and private working material.

3. Scans the candidate for common secret patterns.

4. Runs the public test and documentation checks.

5. Produces a release manifest with file hashes.

Only after human review should that directory be initialized as a new Git repository and pushed once. That keeps private commits and messages out of the public repository.

Boundaries and next evidence

TARA does not yet claim:

• continuous source retrieval and approval;

• comprehensive legal coverage;

• independent tax or legal sign-off;

• production authentication, tenancy, encryption, or retention controls;

• regulator acceptance of a submitted artefact; or

• external traction or a signed design partner.

Presentation guidance

Lead with the concrete failure: an internally consistent 41% calculation is wrong for a 2026 event. Show the AI/MCP trace first to establish the agent story, then show What changed, What applies, the Action plan, the evidence that Needs correction, the evidence that is Accepted, and the Proof record. Explain the deterministic safety boundary in one sentence: the model chooses tools; TARA decides dates, calculations, and evidence outcomes.

License

TARA is released under the MIT License. The regulatory content and demo outputs remain a bounded prototype and do not constitute legal, tax, immigration, or filing advice.