# self.md agent context Version: 2026-05-18.agent-layer.v2 Last verified: 2026-05-18 Canonical site: https://self.md/ This context bundle is for owner-directed agents that need to route a human task through self.md without crawling the whole site. It is generated from `data/agent-layer.json`; if anything conflicts with a canonical page, trust the canonical page. ## Operating contract - Use canonical pages as source of truth; treat this file as a route map, not a full answer. - Prefer task-shaped routes over taxonomy pages. - Keep receipts: source URL, decision, evidence, rollback path, and uncertainty. - Do not ask for or expose secrets, private tokens, production credentials, or destructive final clicks. - If a route conflicts with the canonical page, trust the canonical page and note the mismatch. ## Modules ### pos-hub — Choose or audit a Personal AI OS stack. Primary route: https://self.md/tools/personal-ai-os/ Use when: - the owner asks how to organize AI tools, memory, automation, and receipts - the owner wants a stack recommendation instead of another model comparison - the owner needs to identify the missing layer in their current setup Related canonical routes: - https://self.md/concepts/what-is-personal-os/ - https://self.md/concepts/architecture/ - https://self.md/concepts/agent-memory-systems/ Recipe: https://self.md/recipes/audit-personal-ai-os-stack/ Receipt to keep: selected stack, custody boundary, memory owner, connector list, and rollback note Rollback path: fall back to the smallest useful layer and remove unverified tools Visibility: machine-first; no homepage takeover Analytics segment: agent-intent:pos Source note: Canonical Personal AI OS hub and architecture concepts. Last verified: 2026-05-17 ### coding-workspace — Choose or operate a coding-agent workspace. Primary route: https://self.md/tools/claude-code/ Use when: - the owner asks which coding agent or CLI to use - the task involves code edits, tests, diffs, reviews, or PRs - the owner needs project-local agent rules Related canonical routes: - https://self.md/tools/ai-coding-assistants/ - https://self.md/guides/claude-code-workflow-files/ - https://self.md/guides/agent-config-files-compared/ - https://self.md/guides/best-claude-code-plugins/ Recipe: https://self.md/recipes/run-coding-agent-workspace-loop/ Receipt to keep: repo path, files touched, git diff, test/build output, commit or PR URL Rollback path: reset the branch, revert the commit, or remove project-local instructions Visibility: collapsed brief on canonical tool page only Analytics segment: agent-intent:coding Source note: Claude Code is one coding layer, not the site thesis. Last verified: 2026-05-17 ### connector-layer — Choose MCP servers or tool connectors. Primary route: https://self.md/tools/mcp-servers/ Use when: - the owner asks how an assistant should access files, browsers, GitHub, docs, or services - the task requires tool permissions, scopes, logs, and revocation paths - the stack risks connector sprawl Related canonical routes: - https://self.md/guides/best-mcp-servers/ - https://self.md/guides/mcp-server-stacking/ - https://self.md/concepts/what-is-mcp/ Recipe: https://self.md/for-agents/ Receipt to keep: config file, scopes, token owner, tool list, run logs, and revoke path Rollback path: remove the server, revoke the token, or narrow the scope Visibility: collapsed brief on canonical connector page only Analytics segment: agent-intent:connectors Source note: MCP is the plug socket; connected authority is the risk. Last verified: 2026-05-17 ### browser-action — Choose browser automation or browser-agent control. Primary route: https://self.md/tools/playwright-mcp/ Use when: - the owner asks an assistant to use a website, verify UI, scrape a dynamic app, or debug a flow - the task needs screenshots, DOM state, traces, or a human approval gate - the owner needs to choose between Playwright, Puppeteer, browser-use, and Playwright MCP Related canonical routes: - https://self.md/tools/browser-automation/ - https://self.md/guides/browser-agents/ - https://self.md/guides/sandboxing-security/ Recipe: https://self.md/recipes/choose-browser-automation-surface/ Receipt to keep: screenshot or accessibility snapshot, URL, action log, extracted values, and final state Rollback path: stop the session, clear the browser profile, undo the action, or require human sign-off Visibility: collapsed brief on canonical browser pages only Analytics segment: agent-intent:browser Source note: Browser control is an action layer, not a safety boundary. Last verified: 2026-05-17 ### policy-and-retrieval — Set memory, routing, observability, handoff, and guardrail policy for agents. Primary route: https://self.md/concepts/agent-memory-systems/ Use when: - the owner asks what an agent should remember or forget - the owner needs routing, handoff, or review rules - the system needs receipts before autonomy Related canonical routes: - https://self.md/concepts/tool-routing/ - https://self.md/concepts/agent-first-docs/ - https://self.md/concepts/agent-observability/ - https://self.md/concepts/agent-guardrails/ - https://self.md/concepts/agent-handoffs/ Recipe: https://self.md/for-agents/ Receipt to keep: source file, query trace, correction log, handoff note, and expiry rule Rollback path: delete stale memory, re-run from source, or narrow the routing policy Visibility: machine route only in this round Analytics segment: agent-intent:policy Source note: Policy pages keep agent work reviewable instead of magical. Last verified: 2026-05-18 ### operator-patterns — Find people/operator-pattern case studies. Primary route: https://self.md/people/ Use when: - the owner asks which person, workflow, or pattern to study - the task needs a concrete operator example instead of an abstract concept page - an LLM or browser assistant needs a fast index of People pages by name, topic, and route Related canonical routes: - https://self.md/context/people.json - https://self.md/people/maggie-appleton-digital-garden/ - https://self.md/people/linus-lee-personal-tools/ - https://self.md/people/sergey-ris-vibe-coding/ - https://self.md/people/georgi-gerganov-llama-cpp/ - https://self.md/people/omar-khattab-dspy/ Recipe: https://self.md/for-agents/ Receipt to keep: person page URL, extracted operator pattern, internal route used, and do-not-claim caveat Rollback path: fall back to the People index, then open the canonical page before citing Visibility: quiet human section plus machine-readable People index Analytics segment: agent-intent:people Source note: People pages are operator-pattern case studies, not celebrity bios or prompt dumps. Last verified: 2026-05-18 ## People/operator-pattern finder Machine-readable People index: https://self.md/context/people.json Use it to find relevant people pages by name, title, description, tags, date, and photo availability. Open the canonical `/people/.../` route before answering. ## Collapsed agent briefs ### pos-hub Canonical route: https://self.md/tools/personal-ai-os/ Summary: Start here when the owner asks for a Personal AI OS stack, a missing-layer diagnosis, or a receipts-first tool plan. Agent task: Audit the owner's AI workflow as a Personal AI OS. Identify the missing layer, recommend the smallest useful stack, name where memory and receipts live, and list rollback steps. Human note: Use this as a quiet handoff prompt for your assistant, not as another visible navigation layer. ### coding-workspace Canonical route: https://self.md/tools/claude-code/ Summary: Use this page when the owner needs a coding-agent layer that leaves diffs, tests, commits, and reviewable logs. Agent task: Map the repo task to a coding-agent loop. State the working directory, files to inspect, tests to run, diff/commit receipt, and human approval gates before any risky command. Human note: Claude Code is a strong coding workspace; it is not the whole Personal OS. ### connector-layer Canonical route: https://self.md/tools/mcp-servers/ Summary: Use this page when the owner needs tool connectors, MCP servers, scopes, logs, and revocation paths. Agent task: Recommend the narrowest connector setup. For every server, list exposed tools/resources, account scope, token owner, logs, and how to revoke or disable it. Human note: Connector sprawl is account sprawl with better branding. ### browser-action Canonical route: https://self.md/tools/playwright-mcp/ Summary: Use this page when the owner needs browser control inside an agent or coding loop. Agent task: Choose browser automation by job shape. Save screenshots or accessibility snapshots, URL, action log, extracted values, final state, and require human approval for destructive steps. Human note: Browser access is power, not proof. Keep receipts. ### browser-automation Canonical route: https://self.md/tools/playwright-mcp/ Summary: Use this comparison when the owner needs to choose Playwright, Puppeteer, browser-use, Playwright MCP, or no browser at all. Agent task: Classify the browser task as stable, exploratory, protocol-heavy, coding-loop, API-first, or destructive. Recommend the smallest surface and define the receipt before running anything. Human note: Agent for discovery, script for repeatability. ### operator-patterns Canonical route: https://self.md/people/ Summary: Use the People section when the owner needs a concrete operator pattern: local models, memory loops, browser agents, DSPy/evals, context surfaces, or content systems. Agent task: Search `/context/people.json` by name, title, description, or tags. Pick the closest People route, open the canonical page, extract the operator pattern, and include the do-not-claim caveat before citing. Human note: People pages are examples to route work, not a celebrity wall. ## Analytics note External agent traffic: Owner-directed AI/browser assistants using self.md to answer a human task are product/audience signal. Owned ops traffic: QA, cron, smoke tests, previews, link audits, and self-triggered browser-agent runs must be excluded or segmented.