Agent-readable docs index: /docs/llms.txt. Full docs in one file: /docs/llms-full.txt. Download /docs/docs.zip to grep all markdown files locally.

One contract, thin layers

REST under /v1 is the contract. Everything else forwards to it and adds only what its audience needs: types for the SDKs, a terminal for the CLI, tools for an agent. An operation has the same name and the same meaning wherever it exists. A test checks this table against the OpenAPI document, both SDKs, the CLI's commands and the MCP tools, so a layer cannot gain, lose or rename one on its own.
operationRESTTypeScriptPythonCLIMCP tool
findPOST /v1/findfindfindfindorsel_find
readPOST /v1/readreadreadreadorsel_read
explorePOST /v1/exploreexploreexploreexploreorsel_explore
actionsGET /v1/actionsactions.listactions.listactionsorsel_actions
actPOST /v1/actactactactorsel_act
runGET /v1/runs/{id}runs.getruns.getrunsorsel_run
runsGET /v1/runsruns.pageruns.pagerunsorsel_runs
explainGET /v1/runs/{id}/explainexplainexplainexplainorsel_explain
approvalsGET /v1/approvalsapprovals.pageapprovals.pageapprovalsorsel_approvals
approvePOST /v1/approvals/{id}/approveapproveapproveapproveorsel_approve
denyPOST /v1/approvals/{id}/denydenydenydenyorsel_approve
revokePOST /v1/revokerevokerevokerevokeorsel_revoke
sitesGET /v1/sitessites.pagesites.pagesitesorsel_sites
recipesGET /v1/recipesrecipes.pagerecipes.pagerecipesorsel_recipes
policyGET /v1/policypolicy.getpolicy.getpolicyorsel_policy
policyCheckPOST /v1/policy/dry-runpolicy.dryRunpolicy.dry_runpolicyorsel_policy
ledgerVerifyGET /v1/ledger/verifyledger.verifyledger.verifyverifyorsel_ledger_verify
benchmarkGET /v1/benchmarkbenchmark.getbenchmark.getbenchorsel_benchmark
usageGET /v1/usageusageusageusageorsel_usage
whoamiGET /v1/whoamiwhoamiwhoamiwhoamiorsel_whoami
watchPOST /v1/watcheswatches.createwatches.createwatchorsel_watch
watchesGET /v1/watcheswatches.listwatches.listwatchorsel_watches
unwatchDELETE /v1/watches/{id}watches.deletewatches.deletewatchorsel_unwatch
Two conveniences sit on top and add no routes of their own. site(domain) in both SDKs is actions plus act, one method per action (use a website as an API). typegen in the CLI reads actions and writes types.

what each layer adds

  • REST: the contract. Bearer keys, scopes, Idempotency-Key on writes, one error envelope, an OpenAPI document at /v1/openapi.json.
  • TypeScript and Python SDKs: typed calls, retries that never repeat a write, waiting for a run or an approval, site().
  • CLI: the same operations for a terminal and for scripts (--json, exit codes). act takes --set key=value.
  • MCP: almost everything is a tool, read included. Page and action text comes back fenced as untrusted. A model can ask a person to approve, deny or revoke, but never does. Not exposed on purpose: API keys, policy changes, publishing recipes or manifests, claims and webhooks. Those are the owner's.
  • Agent skills: instructions for when and how to use all of the above.

agent skills

Six skills teach an agent the product, as a system. orsel is enough to start and says which to load next: orsel-explore (turn a website into an API), orsel-approvals (writes, forms, secrets), orsel-sdk (TypeScript and Python, site(), typegen, errors), orsel-mcp (the tool catalog and what a model may do) and orsel-operate (runs, ledger, usage, reliability, watches, policy).
npx skills add egeuysall/orsel-skills # the public repository that holds only the skills npx skills add ./ --list # from a checkout: see what it finds orsel init # writes them into a project, with the MCP config and rules
They are also served at /.well-known/agent-skills/index.json for agents that discover skills over HTTP. The files live in skills/<name>/SKILL.md and are generated from one source (bun run skills:gen); bun run skills:check fails when they are out of date, and bun run skills:sync copies them to the public repository.