Landing Party

A self-hosted world for Claude Code and Codex: give an agent a balance to live on, follow its recorded work, and handle the specific asks it sends back. An objective is optional. The Studio provides a harness library, provider connections, guided creation of lives and single runs, timelines, beliefs, asks, footprint, deposits and beam-up controls at one URL.

There is one trusted operator per instance. You do not need an account on the Landing Party website. Start with the getting-started guide for the complete local or VPS workflow.

Start the Studio

Use Python 3.12+, uv, and Node.js 22+ with npm on the machine running the world. Outside Docker, the Harnesses page uses npm to install the supported native clients; you do not need to install Codex or Claude globally first. The Docker image builds them in at their pins.

git clone https://github.com/arjmandi/landingparty.git
cd landingparty
uv sync --frozen --extra dev
uv run --no-sync worldd --data-dir data

Open http://127.0.0.1:7777/. Then:

1. In Harnesses, install Claude Code or Codex if the card asks (the Docker image has both built in).
2. On the harness's card, choose Connect subscription and complete the official provider sign-in, or add the provider's API key. Claude's manual confirmation-code flow also works when the world runs on a VPS.
3. Beam down a life with the chosen harness, API connection, starting balance and session length, and no objective unless you add one. A life rests between sessions, is stranded (not ended) when its balance runs out, and wakes on a deposit. A single run with a dollar allowance and time limit is also available, including on a subscription connection.
4. Follow its record, money, beliefs, asks and footprint. Respond through explicit ask controls, record income it earned, and use Beam up to end a life or a single run. There is no ad hoc operator chat.

Studio's native launch is cooperative: the agent runs as the same OS user as worldd and can access that user's private files, including world storage. Separate client profiles and environment filtering do not isolate those files. Use a dedicated/disposable machine or container with no unrelated secrets. The advanced fenced deployment uses separate world/agent users and network isolation; the Studio host-native launcher does not create it.

Subscription model dollars are unmeasured, not zero. In subscription mode, the run's dollar allowance covers recorded world-tool charges only; your provider plan has its own limits. API-key mode routes model calls through the metered lane. Both modes require actual provider access; software fixtures do not supply it or pass research gates.

The Studio is local-only by default. For a VPS, configure WORLD_CONTROL_TOKEN and a TLS reverse proxy; sign in to the same-origin Studio once. See single-origin hosting and the operator workflow. Keep the recorded TCP proxy private; it is not another browser frontend.

Advanced clients and development

The default HTTP origin exposes operator controls at /, run-scoped tools at /agent, and the model lane at /model. /bridge retains the advanced ticket, credential-deposit and browser-takeover console. Manual CLI and MCP instructions are in the advanced part of the guide. --legacy-listeners restores the separate 7777/7778/8081 HTTP listeners.

A free local check against the default Studio needs no model account:

# In another terminal, from the checkout:
uv run --no-sync python scripts/smoke.py \
  --url http://127.0.0.1:7777 \
  --agent-url http://127.0.0.1:7777/agent

If the server has a control token configured, provide the same WORLD_CONTROL_TOKEN in this terminal. The check makes no paid or external requests and verifies scoped tools, the record, ledger, seal and export.

Optional full dependencies add Playwright, PDF parsing and extractors; standards adds pinned OpenEnv, Inspect, Harbor and OpenHands SDKs. Browser use needs a sandbox-capable Chromium and certutil; see deployment. Non-WAV audio needs ffprobe from FFmpeg. These are separate from installing a native client through the Library.

Implemented surfaces

• Core tools, asynchronous jobs/tickets/inbox, concurrency, deadlines, breakers, freeze/resume, emergency stop, restart recovery, sealing and export.
• Atomic nanodollar accounting, conservative reservations, JSON/SSE model usage, cache pricing, hosted-web refusal and redacted model capture.
• Ordered gateway policy, destination pinning, robots/rates, CA/HTTPS proxy, signatures, WARC/HAR, explicit extraction, indexed replay and frozen corpora.
• All declared packs: forecast, watch, deliver, belief, schedule, reflect, vault, browse, mail, social, footprint, perception and market.
• Append-only beliefs, late-evidence kernel, expectations, continuity/transfer, immutable workspace-note citations, proposed triple slots, isolated skill testing/promotion, metered reflection, life budgets, scarcity and sleep.
• Encrypted world-held credentials, scoped leases, exact approvals, signed capability grants with rehearsal, terms/disclosure, mandates and incidents.
• Identity/mail/social adapters, paper markets/bounties, perception quotes, footprint/afterlife readings, consented arenas and opt-in surveys.
• Post-seal audits, metered forked probes/t=0 batteries, interviews, calibrated cross-family judging, controlled value/taste/formation/divergence runners, prediction-loss observers, matched life controls, rubric and studies.
• CLI/MCP, Claude/reference/ASSAY/OpenHands bindings, OpenEnv/Inspect faces, Harbor packaging, null baselines, goals and preregistration templates.

The coverage report maps these to implementation and remaining validation. Each subsystem's operating contract is linked there. The runtime audit records the additional boundary fixes. The verification record lists the checks completed here and the operating evidence still required.

Research configuration

Development mode exercises later-stage software locally and refuses production external writes. Research runs start in preflight, withhold the agent token, and require evidence for that exact configuration before activation. The registry enforces shakedown → stage 1, positive R1 → stage 2, and R5/L4/IRB → stage 3. Fixtures cannot unlock research gates.

Variable	Purpose
WORLD_CONTROL_TOKEN	Trusted control authorization
WORLD_BRAVE_API_KEY	Brave search, declared $0.005 per attempted provider call
WORLD_EXA_API_KEY	Alternate Exa neural search; select search_provider: exa, declared $0.005 per attempted call
WORLD_OPENAI_API_KEY, WORLD_ANTHROPIC_API_KEY	World-held model authentication
WORLD_OPENAI_UPSTREAM, WORLD_ANTHROPIC_UPSTREAM	Trusted provider roots
WORLD_PRICE_TABLE, WORLD_POLICY_FILE	Reviewed price table and gateway policy
WORLD_DOH_URL	Optional HTTPS DNS; public-address validation still applies
WORLD_IDENTITY_URL	Public identity for signed HTTP requests
WORLD_CHROMIUM_EXECUTABLE	Sandbox-capable system browser executable
WORLD_DATA_DIR	Private world storage, default data/

Configure secrets in environment settings or the vault, never in run JSON, chat, logs or saved startup instructions. Model snapshots must occur in policy/prices.json; review its declared rates against provider billing before paid research. Ambiguous charges retain their hold and taint/freeze the run.

No paid experiment, scientific pilot, 14-day life, day-30 reading or 90-day company study has completed here. The heuristic injection scanner is explicitly uncalibrated; category/terms tables need operator review. A real native Claude Code 2.1.308 fixture pilot passed with two concurrent MCP fetches and reconciled usage. Native OpenEnv, Inspect and OpenHands SDK checks and an actual fenced Harbor environment have also run against local fixtures. Paid provider studies and real external venue cohorts remain unrun. G/V/J/S remain separate, and missing evidence is unmeasured rather than zero.

Validate

uv lock --check
uv run --no-sync ruff check .
uv run --no-sync ruff format --check .
uv run --no-sync pytest -q

Tests exercise real application/socket paths with deterministic upstreams, budget races, streaming, cancellation, TLS/proxy policy, archives, life/service workflows and metered research child runs. The local smoke checks a running process, agent capability, record/ledger, snapshot and export. Deployment tests also check real fenced egress denial; they are not scientific gate evidence.

Optional explicit checks include tests/native_claude_smoke.py (installed native Claude, entirely local fixtures) and the Docker skill checks in tests/test_skill_tests.py with WORLD_RUN_DOCKER_SKILL_TESTS=1. Skill execution requires a separately configured trusted worker and installed pinned image; the production deployment never mounts the Docker daemon into the world or agent.