Make Landing work for you¶
You choose the model, environment, instructions, tools, and acceptance checks. Start with the built-in commands and adapt the parts your team needs. CLI, CI, and HTTP calls share the same action contract.
Choose the model and environment¶
Configure the model where work executes:
export LANDING_MODEL="openai:gpt-4.1"
export LANDING_API_KEY="your-provider-api-key"
You can save settings in ~/.landing/config.yml; see Configuration for endpoints and precedence. Model requests go to the selected provider, whose data handling applies.
Prepare checkouts, language runtimes, dependencies, credentials, and baseline data using your existing setup process. A remote client uses the server's prepared environment. Landing does not install project tools or skills for you.
Provide project guidance¶
Landing reads the selected workspace's root AGENTS.md. Its guidance asks the agent to read more specific instructions along affected paths. Put durable project conventions there, and give task-specific requirements in the instruction or input files.
# Project instructions
Use python acceptance.py to verify CLI behavior. Compare benchmarks with the same workload and environment. When investigating a deployed failure, record the deployed revision and the relevant logs.
Instructions guide the agent; they do not grant permissions or enforce an operating-system boundary.
Prepare skills¶
Skills live in a named directory with a SKILL.md containing name and description YAML front matter. Landing discovers the workspace's .agents/skills, additional trusted roots, ~/.agents/skills, and its bundled mode skills. Each mode uses its permitted landing-issuer, landing-fixer, landing-gatekeeper, landing-explainer, or landing-bumper method automatically. To replace a built-in method, prepare a same-named skill in your project or another trusted root. The agent can load a permitted skill through its skill tool; $skill-name includes it directly in an instruction.
landing explain 'Use $deployment-check to explain the failed deployment.' --workspace ./candidate --input deployment.log
For a skill repository on GitHub, prepare it with your normal gh login:
gh repo clone example/team-skills ~/.local/share/landing/team-skills
landing --skill-dir ~/.local/share/landing/team-skills/.agents/skills explain 'Use $deployment-check to explain the failure.' --workspace ./candidate --input deployment.log
Configure roots on the executing host. Use skill_dirs or LANDING_SKILL_DIRS for saved settings and repeatable global --skill-dir options for individual calls. Skills configuration defines discovery precedence. Skills do not install plugins or expand configured tools.
Select tools and checks¶
Each mode can independently select its allowed or excluded tools and skills through mode capabilities. Enforce authorization in those tools and the execution environment; a shell tool can expose more than its name suggests.
Use your existing checks for the behavior you need to establish:
landing fix "Resolve the reported compatibility problem." --input acceptance.txt --check "make acceptance" --check "make benchmark"
Checks execute in the task workspace. Exit status establishes pass or failure, so a benchmark command must enforce your performance criterion rather than merely print a report.
Connect and extend¶
Supply focused evidence with --input, use the prepared gh CLI for GitHub, or translate another platform's events into the HTTP contract. Provider signatures belong at that translation boundary. Use a service for shared callers.
Landing's Apache-2.0 source lets you change prompts, runtime behavior, and adapters. Use the CLI, GitHub integration, or HTTP service to delegate work. For an embedded agent, use Bub's SDK. Actions and model history stay in SQLite; workspace files remain separate. Choose retention and recovery for both.