Gridzen Developer Kit 0.6.0

Research for 198 countries/territories, a verification integration planner, five synthetic outcomes, an HTTP SDK, local MCP server and six Skills. No live provider is enabled. No real person is verified. The research snapshot is dated 2026-10-07, with source links and file hashes.

中文指南 · Guía en español

Try it

Public playground: https://gridzen.ai/developers/

Download gridzen-developer-kit.zip, extract it, then:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install './gridzen-developer-kit[mcp]'
gridzen coverage --country ID --capability bank_account_match
gridzen plan --country ID --event payout
gridzen simulate --country ID --capability bank_account_match --scenario timeout
python gridzen-developer-kit/examples/payout.py

Python 3.11+. CLI and MCP use the bundled offline catalog and fixtures; the Python HTTP client/example calls the public sandbox. Dependencies are downloaded only during installation. API reference: https://gridzen.ai/developers/api/openapi.json .

MCP

Register a local stdio server in your MCP client. Use the absolute path to the installed executable; configuration syntax varies by client. A conventional mcpServers configuration is:

{"mcpServers":{"gridzen":{"command":"/absolute/path/to/.venv/bin/gridzen-mcp","args":[]}}}

Tools: get_coverage, plan_verification, create_sandbox_verification, get_sandbox_verification, explain_result. All operate on local research or synthetic fixtures, with read-only/idempotent annotations. Public research/sandbox MCP endpoint: https://gridzen.ai/developers/mcp (Streamable HTTP, no account or API key). It has the same five tools and zero live providers. Remote usage sends the documented arguments to Gridzen; local stdio is offline by default. Limits: 120 requests/minute per client address and 600/minute overall, returning HTTP 429 with Retry-After.

Skills

Copy the individual directories from skills/ into your compatible agent's project or user skill directory. Each entrypoint is SKILL.md; preserve its references/ directory. No marketplace listing or automatic registration is implied.

HTTP SDK

from gridzen_developer.client import Gridzen
client = Gridzen()
plan = client.plan('ID', 'payout')
result = client.simulate('ID', 'bank_account_match', 'not_found')
assert result['status'] == 'inconclusive'
assert result['verified'] is False

Only documented country/event/capability/scenario fields are accepted. Never send identity numbers, bank accounts, documents or credentials to the sandbox. Simulation IDs encode fixtures and are not unique transaction identifiers. A synthetic match cannot authorize real users or payments. Research evidence is distinct from Gridzen availability or commercial permission.

Repository checkout: local service and checks

python -m pip install '.[test,server]'
python -m pytest -q
python scripts/check_mcp.py
uvicorn gridzen_developer.server:app --host 127.0.0.1 --port 8030

Production uses an independent container at localhost:8030 proxied only under /developers/. See deploy/compose.yml and deploy/developers.conf. Removing that proxy file and stopping this container rolls back the preview; the existing API and Agent Company stacks are separate.

Repository checkout: build distributions

python scripts/build_downloads.py creates allowlisted developer-kit and Skills ZIPs; it excludes tests, deployment files, private credentials and itself from recursive download packaging. python scripts/build_catalog.py <repo>/docs/research/identity-verification/2026-10-07 refreshes the public research projection. Main repository report files remain the provenance source.

Public distribution

Source: https://github.com/immurray/gridzen-developer-kit . Install the Skills:

npx skills add immurray/gridzen-developer-kit

Install the tagged Python/MCP release from GitHub (Python 3.11+ and Git):

python -m pip install "gridzen-developer-kit[mcp] @ git+https://github.com/immurray/gridzen-developer-kit.git@v0.6.0"

PyPI 0.6.0 is published and verified. Install with python -m pip install gridzen-developer-kit. See the release tracker for publication and installation evidence.

Remote client configuration (conventional JSON; client syntax may vary):

{"mcpServers":{"gridzen":{"type":"http","url":"https://gridzen.ai/developers/mcp"}}}

For Claude Code: claude mcp add --transport http gridzen https://gridzen.ai/developers/mcp. For Cursor: add the remote URL in its MCP settings. For portable Agent Plugins, use the packaged plugin.json and mcp.json; the portable transport identifier is streamable-http. Marketplace admission is separate from compatibility.

Original code and Skills: MIT. Research provenance and third-party rights: NOTICE.md. Support: open@gridzen.ai .

Marketplace availability

Source, release downloads, the official MCP Registry entry and six Skills.sh pages are public. Docker catalog submission is awaiting review. See the publication ledger for exact status and remaining platform steps.

Six bundled Skills (0.6.0)

python -m pip install gridzen-developer-kit installs the CLI, stdio MCP dependencies and all six Skill directories. Version 0.6.0 is published on PyPI; clean installation, CLI, stdio MCP and all six bundled Skill directories were verified. No extra MCP dependency installation is required.

gridzen coverage --country MX
gridzen skills
gridzen-mcp

Connect gridzen-mcp as your client's stdio server, or use the remote endpoint https://gridzen.ai/developers/mcp. gridzen skills prints the installed directories; copy a complete directory into your assistant's configured Skills directory to activate it. Installing a wheel does not automatically activate a client's Skills.

The existing select/integrate/explain Skills remain available. Added offline workflows, version 1.0.0: mexico-pilot-scoper, payout-policy-designer, and provider-rights-readiness. Each includes MIT, README, complete response fixtures and agent acceptance prompts. They make no third-party calls and grant no real verification, legal approval or regulatory conclusion.

Configure your assistant (v0.6.0)

python3 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade gridzen-developer-kit
gridzen setup --client all --project .
gridzen setup --client all --project . --apply

Use --client for Codex, Claude Code/Desktop, Cursor, VS Code/Copilot, Copilot CLI, Gemini CLI, Cline, Roo Code or OpenCode. Setup previews by default, merges existing unrelated settings, backs up changed configurations and refuses differing Gridzen entries or Skill files. No global client settings are modified. Native clients require workspace trust; Desktop and Cline require manual MCP import. Desktop produces six separately uploadable ZIPs. See client-specific instructions and actual compatibility evidence. Configuration preparation is not model acceptance or real verification.

Hosted MCP aggregate attribution

Hosted calls count successful tool invocations, not customers. Existing totals remain unchanged. The hosted service records UTC day, tool name, one source category and a count; it never persists request arguments or client addresses. Historical and unlabelled calls remain unknown. Local stdio is offline by default.

Explicit acceptance requests to the loopback-published service may set X-Gridzen-Purpose: internal_test. Only configured local peers with no forwarded client headers are classified as internal_test. The public reverse proxy must overwrite X-Real-IP; publicly supplied labels are only self_reported_test. An unlabelled public request is not an authenticated external customer. The GRIDZEN_MCP_LOCAL_TEST_PEERS deployment setting must match the container gateway and must never contain public client addresses. Source counts use the existing 90-day retention and fail-open behavior. This hosted change does not publish a new PyPI version or enable any live provider.

Customer tasks and voluntary feedback (0.6.0)

gridzen plan --country US --event onboarding --task signup_phone
gridzen plan --country US --event onboarding --task identity_onboarding
gridzen plan --country US --event onboarding --task identity_onboarding --stage production
python examples/customer_tasks.py
gridzen feedback --task signup_phone --country US --outcome blocked --blocker missing_workflow_step --output task-summary.json
# Only after choosing to share:
gridzen feedback --input task-summary.json --share

Phone intelligence is not OTP or ownership verification. Identity fixtures are not hosted document/liveness sessions. Payout matching remains a prototype with no real enabled route. Provider intake Skills are for provider operations/BYO, not a universal customer gate.

Remote APIs count fixed task/capability/country/stage/error categories for 90 days, including technical failures. They store no arguments, identity data, credentials, raw chats or per-user tracking. Local feedback is off by default. After explicit opt-in, CLI and stdio MCP automatically upload category-only events. Reading static Skills remains unobservable. Optional feedback requires explicit consent, contains only fixed categories, and is unverified self-reported evidence. A SHA-256 hash of its random receipt ID is retained 30 days for 30-day retry deduplication; the ID itself is not stored. Counters are capped at 10,000 requests/day and 500 summaries/day per database; missing history or offline usage is unknown. Requests are not unique customers.

Optional automatic local feedback (0.6.0)

gridzen telemetry status
# Run only after reading and agreeing to category-only automatic sharing:
gridzen telemetry enable --consent
gridzen plan --country US --event onboarding --task signup_phone
# Send one queued event per invocation after reconnecting:
gridzen telemetry flush
# Disable future uploads and clear the pending queue:
gridzen telemetry disable

No network calls occur from local CLI/MCP when disabled. Consent is saved per OS user, not silently added to a project. Events contain only documented categories, ISO country, and a random per-event receipt, with no conversation, identity fields, credentials, user or project IDs. Feedback failures do not fail tool execution. The local queue is capped at 1,000 events/30 days; reconnect or flush retries one queued event, with a one-second HTTP timeout. Uploads are self-reported usage, not authenticated customers or completed verification. No background daemon is installed. Merely reading a Skill is unobservable; after consent, an Agent can explicitly record a fixed-category task summary.