dcbuilder-cli - SKILL.md Agent Skill

name: dcbuilder-cli description: Use when an agent needs to operate the dcbuilder CLI or local MCP server to read dcbuilder.dev news, jobs, candidates, run allowlisted queries, submit inbox messages/jobs/candidates, manage inbox review actions, create submit invites, refresh search indexes, or fetch the agent API schema.

dcbuilder CLI

Use this skill whenever the user asks an agent to work through the dcbuilder CLI or MCP server. The backend API is the source of truth; the CLI must not query Postgres directly.

Environment

Required for authenticated commands:

export DCBUILDER_API_URL=https://dcbuilder.dev
export DCBUILDER_API_TOKEN=...

Do not store DCBUILDER_API_TOKEN in repo files, config files, logs, or examples. Use environment variables only. Public submit message may be allowed without a token, but authenticated usage is preferred for trusted agents.

Default output is JSON. Use --format table or --format markdown only for human-readable summaries.

Read Commands

Fetch OpenAPI/schema metadata:

dcbuilder schema

Read recent news:

dcbuilder news --since 7d --limit 20
dcbuilder news --q "MegaETH" --format table

Read jobs:

dcbuilder jobs --remote --limit 25
dcbuilder jobs --company "ExampleCo" --location "Remote"
dcbuilder jobs --tag infra --tag ai

Read candidates:

dcbuilder candidates --availability open --location "Europe"
dcbuilder candidates --q "founding engineer"

Sensitive candidate fields are redacted unless the token has elevated scope such as agent:sensitive, admin:read, or *.

Query Command

Use structured filters as the canonical contract:

dcbuilder query --json '{"table":"jobs","filters":{"remote":true,"limit":10}}'

Natural-language query text is only client-side convenience. The CLI parses it into deterministic filters before calling the API:

dcbuilder query "remote AI infra jobs from the last week"

Never expose raw SQL as a CLI/MCP interface.

Submit Commands

Submissions create inbox items for review; they do not publish live jobs or candidates by themselves.

dcbuilder submit message --json '{"subject":"Scout report","body":"Found 3 relevant jobs.","source":"agent","priority":"normal"}'
dcbuilder submit job --json '{"title":"Founding Engineer","company":"ExampleCo","link":"https://example.com/jobs/1","location":"Remote"}'
dcbuilder submit candidate --json '{"name":"Ada Example","headline":"AI infra engineer","location":"Berlin","skills":["Bun","MCP","Postgres"]}'

Use submit message when reporting findings, asking for approval, or requesting help from Daniel. Prefer messages over direct approval unless the user explicitly asked the agent to approve or publish.

Inbox Review Commands

Admin-scoped tokens can review and manage inbox submissions:

dcbuilder inbox list --status pending --limit 20
dcbuilder inbox show <id>
dcbuilder inbox comment <id> "Looks good; please add salary range."
dcbuilder inbox approve <id> --json '{"note":"Approved by trusted agent"}'
dcbuilder inbox reject <id> --json '{"reason":"Duplicate submission"}'

Approval is create-only for job/candidate records. Do not approve, reject, or comment unless the user's instruction clearly authorizes that action.

Admin Commands

Create scoped submit invites:

dcbuilder invites create --json '{"allowedKinds":["job","candidate","message"],"submitterName":"Partner","submitterEmail":"partner@example.com","expiresInDays":30,"maxUses":5}'

The returned invite token is a secret. Do not print it in final answers unless the user explicitly asked for the token value.

Refresh search snapshots / semantic index metadata:

dcbuilder search refresh --json '{"resources":["jobs","candidates","news"],"limit":100}'

If no embedding provider is configured, the backend should fall back cleanly to filter/text search.

MCP Server

Start the local stdio MCP server from the CLI package:

dcbuilder-mcp

The MCP server uses the same DCBUILDER_API_URL and DCBUILDER_API_TOKEN environment variables. It exposes typed tools for:

dcbuilder_schema
reading news, jobs, and candidates
generic allowlisted queries
submitting messages, jobs, and candidates
listing/showing/commenting/approving/rejecting inbox items
creating scoped submit invites
refreshing search indexes

Prefer MCP tools when the calling agent supports MCP natively. Prefer CLI commands when producing a reproducible shell transcript or operating in a plain terminal.

Agent Workflow Pattern

Read with news, jobs, candidates, or query.
Analyze locally; do not send free-form prompts to the API as the canonical query contract.
Submit a concise inbox message with findings and recommended next action.
Only use inbox approval actions when the user explicitly delegates approval.

Example scout flow:

dcbuilder news --since 7d --limit 20 > /tmp/dcbuilder-news.json
dcbuilder jobs --remote --limit 25 > /tmp/dcbuilder-jobs.json
dcbuilder candidates --limit 25 > /tmp/dcbuilder-candidates.json
dcbuilder submit message --json '{"subject":"Weekly opportunity scout","body":"Summary with recommended follow-ups...","source":"dcbuilder-cli","priority":"normal"}'