Sourcegraph Deep Search

What it does

Deep Search is an agentic layer over Sourcegraph code search and its code graph. It answers natural-language questions spanning repositories, symbols, and relationships. The package includes code search, Batch Changes, Insights, Monitoring, MCP, APIs, and CLI access; Claude Code, Cursor, Codex, and Amp are named clients. The practical question is not whether a polished demonstration works once, but whether it produces repeatable results with representative data, clear ownership, and predictable cost. Relevant comparisons are Cursor Agent, GitHub Copilot Agents, Phind. These products overlap with only part of the workflow, so compare actual inputs, outputs, integrations, and review effort—not the number of labels on a feature page.

Pricing and operating cost

The first-party pricing page starts at $16,000 with a minimum annual contract. AI credits are included and pooled organization-wide without monthly expiry; purchased buckets roll over on renewal. It lists single-tenant cloud, self-hosting, 24×5 support, a success manager, and optional premium support. Prices can change after publication. Confirm currency, taxes, annual commitment, included usage, expiration, overages, rate limits, cancellation, refunds, support, and minimums at checkout. Model both a normal and peak month. Subscription price omits operator review, failed runs, storage, integration maintenance, and downstream infrastructure.

Strengths and drawbacks

The strongest reasons to shortlist it are: Surfaces cross-repository relationships; Serves humans and agents through several interfaces; Enterprise deployment options. Those advantages matter only when they remove a measured bottleneck. Assign an owner, define acceptable output, and retain original evidence or artifacts so results remain auditable.

The material drawbacks are: $16,000 minimum excludes small teams; AI credits add variable cost; Results need runtime verification. Do not hide those behind a generic warning that AI can err. Use failure-specific controls: least-privilege access, human approval for consequential output, spending alerts, versioning, logs, deletion procedures, and a fallback. Treat vendor performance claims as hypotheses until reproduced on the buyer's workload.

Practical use cases

Useful deployments include Incident tracing; Codebase onboarding; Migration impact mapping; Governed MCP context. Begin with one bounded workflow where volume, baseline time, correction rate, and success are measurable. Avoid broad rollout until administrators understand permissions, retention, export, and failure handling. Qualified review remains mandatory for regulated, confidential, customer-facing, legal, financial, medical, or production-code work.

How to evaluate it

Collect 40 known questions from onboarding, incidents, migrations, and security work. Score recall, false leads, citations, permissions, index freshness, and answer time. Test restricted and deleted repositories and ingestion lag after merges. Run long enough to capture difficult cases, not only happy paths. Record total cost per accepted result and the percentage needing correction. Review security documentation, subprocessors, model-training policy, data residency, incident response, portability, and contractual liability. Purchase only when measured time or quality gains exceed subscription, integration, governance, and review costs.