# 17.5 When to write your own connector (/building-ai-workers/connectors-and-mcp/built-in-custom-skill-or-connector) --- type: Document title: "17.5 When to write your own connector" description: "When a connector is worth writing, and when a script or a skill is the cheaper piece." status: stable order: 317.05 ksor: owner: team:panaversity audience: [ public ] approval: by: process:panaversity at: 2026-11-02T18:20:08Z chapter: "17" part: III expert_status: co-build objectives: - { id: CCDV-F.D8.S3, label: direct } - { id: OAI.2.5, label: direct } concepts: [ "17.5" ] last_verified: 2026-11-01 generated: at: 2026-11-02T18:20:08Z by: esl-rewrite/1.2.0+ksor.1 trust_tier: unverified build_id: sha256:e1fbcb7d9a8f8b9b7cb96875022325c02c9ba4be94cef13b301cbd39a9edad85 dirty: true ksor_version: 0.0.60 --- **In everyday life.** For dinner, you can cook with what you have (a built-in tool), follow a recipe card (a skill), or order in (a connector). Ordering in is worth it when the food is out of your reach. To order in, you can pick from a restaurant's menu (a ready-made connector), or hire a cook who makes exactly your dishes (your own connector). Neither is worth it for a sandwich you can make yourself. **What you will learn.** When a job needs a connector, and when writing your own is worth the upkeep. **Why it matters.** A connector is a service someone must keep running, so writing one for a job a script could do wastes time and money. **Two jobs, start to finish.** Ana has two jobs she wants an AI to help with: - **Policy search.** AP clerks use a chat app, the builder uses a coding agent, and finance tries a different AI app. All three need cited answers from the same AP policies. - **The due-by-Friday list.** One person runs it, in one project folder, with one coding agent. A tested script already does it. The first job becomes a tool in a connector Brightline writes. The second stays a script inside a skill. Why the difference? ### How do I decide where a job belongs? Chapter 16 placed each job by asking questions in order, and stopping at the first yes. Must it never happen? A hook or a setting. Is it in another system? A connector. Must a number be exact? Your own tested code. Can the agent already do it? A built-in tool. Is it side work that would crowd the session? A helper agent. Otherwise, a skill. This chapter widens the connector question and adds a second one. ### First question: does the job need a connector at all? Yes, when at least one of these is true: - **It is in another system** the AI cannot reach on its own, such as an accounting system, a database, or files behind a login. - **Several AI apps need it.** One connector serves every AI app that speaks MCP. A script serves one agent in one project. - **A rule must hold whichever app asks,** such as fields that must never leave. The connector applies the rule for every app. The due-by-Friday list meets none of these. It reads only the weekly invoice export, which holds no bank details, in the one project folder its one user works in. So a tested script can read the file directly, and no connector is needed to guard it. A skill says when to run the script. Policy search meets the second: three AI apps need the same answers. ### Second question: a ready-made connector, or your own? Look for a ready-made one first, and use it if it reaches your system and gives only the access you want. Write your own when: - **None knows your system or your work,** such as Brightline's purchase orders and AP policies. - **The ready-made one gives too much,** as Chapter 16's did: it could change files, and it handed over whole files. - **It cannot keep your rules,** such as fields that must never leave, or a password for the business system that the model must never see. ### What does a connector cost? Every tool description is sent to the model each time it is asked anything. A connector is also a running service: it has a web address, an access key, a place it runs online, and someone who answers when it breaks. So when a tested script or a skill does the job, use that. ### What does a connector control, and what does it not? A connector controls what can be asked and what data comes back. It cannot tell whether a person really wants an action done right now. That decision belongs to the AI app's settings and, for anything that changes data or moves money, to a person who approves it ([17.7](untrusted-output.md), Chapter 23). **At work.** Ana and the builder settle both jobs as above. Policy search needs a connector, because three AI apps need it. Chapter 16's ready-made connector could open the policy files, but only as whole files. It could not search them or cite a line. So policy search becomes the `search_policy` tool in Brightline's own connector. The due-by-Friday list stays a tested script in a skill, because a connector would add upkeep without a single new user. **The operating rule:** use a connector when the job is in another system, when several AI apps need it, or when a rule must hold whichever app asks. Use a ready-made one if it fits. Write your own when no ready-made one fits: none knows your system, it gives too much, or it cannot keep your rules. Otherwise, use a skill or a tested script. ![Chapter 16's questions, asked in order. Must it never happen: a hook or setting. Another system, several AI apps, or a rule for every app: a connector, highlighted. Must a number be exact: your own tested code. Can the agent already do it: a built-in tool. Side work that would crowd the session: a helper agent. Otherwise, a skill. Beside the connector, two choices: a ready-made one if it fits, or your own. Write your own when none knows your system, the ready-made one gives too much, or it cannot keep your rules.](img/ladder.png) *Figure 17.10. Inside the connector step there are two choices: use a ready-made connector if one fits, or write your own when none knows your system or keeps your rules.*