mrkeyoor.com_
Tue 11 Aug 13:28 UTC
Self-Hostedevaluationupdated 11 Aug 2026

sourcebot

Sourcebot is a self-hosted search and browsing service for code spread across many repositories and hosting platforms. It also lets people ask cited questions about that code and can expose search, file reading, and code navigation to AI tools through MCP.

Verdict

Sourcebot is an excellent first trial for teams whose code is fragmented across hosts and whose current search tools waste time. The Docker start, polished browsing, Zoekt search, cited answers, and active release cadence make it much more practical than assembling those pieces yourself. Before standardizing on it, price the paid access controls, accept the source-available license, and test indexing correctness and resource use on your real repository set.

Setup4/5Fast Docker start, with real auth and storage work for teams
Docs5/5Detailed deployment, sizing, auth, indexing, model, and MCP guides
Community5/5Frequent releases and busy issue and pull-request activity
Maturity4/5Polished v5 product with scaling and index-correctness caveats

Who it’s for

Engineering teams that need one fast search box across many repositories and code hosts.
Organizations willing to operate a code index so private source stays in their infrastructure.
Developers who want regular expressions, filters, branch indexing, definitions, references, and file browsing in a web UI.
Teams prepared to pay for permission syncing or an MCP context layer for Claude Code, Codex, Cursor, and similar tools.

Who it’s NOT for

Companies that require a permissive open-source license or want to sell a competing hosted service: most code uses FSL-1.1-ALv2, which restricts competing use until that version converts to Apache 2.0 after two years.
Multi-user teams needing least privilege on the free plan: the docs say every signed-in user receives the Owner role, while role management and repository permission syncing are paid.
Operators who need proven horizontal scaling across thousands of repositories: the sizing guide describes a vertically scaled container, and RFC #439 reports web unresponsiveness and out-of-memory failures under heavy indexing.
Auditors relying on exhaustive result counts: issue #504 says search limits can stop Zoekt without telling the UI or logs, producing incomplete results without a warning.
Teams expecting every OpenAI-compatible endpoint to power Ask reliably: issue #1105 documents hanging streams and malformed tool-call handling in that path.
Anyone needing binary-file search: the indexing guide says binary files cannot be indexed.

Setup reality

A small trial is genuinely quick: run the setup CLI or create one JSON configuration, start Docker Compose, and open port 3000. Budget at least 4 GB of memory, plus code-host credentials, persistent storage, and an LLM key if you want Ask. Production requires more care: disk should be two to three times the source size, branch indexing can multiply storage, and larger installs should move Postgres and Redis out of the all-in-one container. Authentication, TLS, backups, telemetry policy, license activation, and repository permissions must be configured deliberately.

Code search is still the strongest pitch

Sourcebot gives a team one place to search code across GitHub, GitLab, Bitbucket, Azure DevOps, Gerrit, Gitea, generic Git hosts, and local repositories. Underneath the interface is Zoekt, a trigram index built for fast regular-expression search. The web UI adds filters, Boolean queries, file browsing, and symbol navigation.

That package solves a mundane but expensive problem. Engineers often know a function, configuration key, or error string exists but not which repository owns it. Sourcebot clones configured repositories, refreshes them in background jobs, and presents one query language across hosts.

Ask Sourcebot builds on that useful base rather than replacing it. A reasoning model receives search, navigation, and file-reading tools, then returns an answer with inline code citations. You choose and fund the provider, with hosted and OpenAI-compatible options available. Quality still depends on the model and indexed evidence, but citations make answers easier to check than generic code chat.

The first deployment is refreshingly small

For a trial, Sourcebot has the right shape. npx setup-sourcebot can generate configuration and Docker Compose files, or an operator can download the compose file, write a short JSON configuration, add secrets, and run one command. The application appears on port 3000. A search-only pilot does not require an LLM provider.

The single container includes the Next.js server, indexing workers, Zoekt, Postgres, Redis with BullMQ, and a persistent cache. That convenience explains the four-gigabyte minimum and need to preserve volumes. Sourcebot recommends external managed Postgres and Redis for more stable deployments, and teams need backups for business data and indexes.

Storage planning matters. The sizing guide suggests disk equal to two or three times the source being indexed. Enabling many branches can triple index storage in the project's testing. Memory directly affects search because the operating-system page cache keeps hot index data available. Recommendations grow from 4 GB of RAM for up to 100 repositories to 64 GB or more beyond 2,000.

Access control changes the price calculation

Self-hosted does not mean every feature is included without a subscription. On the free plan, all signed-in users become Owners and role management is unavailable. Repository permission syncing, which limits search, browsing, Ask, and MCP results to repositories a user can access on the original host, is paid. That boundary matters when teams share one instance but not one source-code permission set.

The Sourcebot MCP server is also marked as paid. It exposes search, file reading, definitions, and references over streamable HTTP to Claude Code, Codex, Cursor, VS Code, and other clients. API-key authorization is available, while OAuth is paid too. This can be a strong company-wide context layer, but it belongs in the subscription comparison.

Telemetry is enabled by default. The project says it sends sanitized usage and performance metadata through PostHog, not code, queries, credentials, or IP addresses. A single environment variable disables it.

The license is source-available, with a timer

Most code uses Functional Source License 1.1 with an Apache 2.0 future license. Internal use, education, research, modification, and redistribution are permitted, but using it to provide a competing commercial product or service is prohibited. Each version becomes available under Apache 2.0 on the second anniversary of its publication.

Enterprise directories use a separate commercial license, and third-party components retain their own terms. This is workable for internal deployment, but it is not equivalent to an immediately permissive open-source project. Vendors and anyone redistributing modifications should review the exact files and version dates.

Search correctness and scale need testing

Open issue #504 is the most important functional caveat. Zoekt has limits on matches and wall time, but Sourcebot does not currently tell the user when those limits stop a query. A search that looks exhaustive may be incomplete. That is dangerous for questions such as whether a pattern exists anywhere or how many instances remain. Audits need another method or an explicit check of limits.

Index health can also be too quiet. Issue #1361 describes ctags failure producing no symbols while the interface reports successful indexing. Issue #1025 reports deleted or renamed branches remaining searchable after a successful sync. Both concern older v4 builds, so reproduce them against v5.1.6, but acceptance tests should still cover symbol failure, branch deletion, and missing shards.

Horizontal scale remains a work in progress. The official sizing guide describes vertical scaling, while an open RFC notes that heavy indexing has made the web application unresponsive or caused out-of-memory errors. A large buyer should test indexing and search together at realistic load.

Health and the decision

The latest release, v5.1.6, shipped on August 10, 2026, and the repository was pushed again on August 11. It included query fixes, private-repository review support, MCP compatibility work, a Zoekt update, and many dependency security upgrades. The 101 open issues and pull requests combined sit beside fast-moving merged work and new contributors, signaling an active product.

Sourcebot deserves a trial when multi-repository search is the actual pain. Its basic experience is focused, its documentation is excellent, and AI features rest on a credible search engine. Access control, MCP, licensing, exhaustive-query behavior, and large-scale operation may change the economics. Test those boundaries with your own code before declaring it the company search layer.

Alternatives

ProjectWhat it isPick it when
OpenGrokA long-running source search and cross-reference engine built in Java.pick this instead when classic code search and history matter more than built-in AI answers or an agent-facing MCP layer.
HoundA lightweight web interface for fast regular-expression searches across repositories.pick this instead when you want a simpler MIT-licensed search service with far less product surface to operate.
LivegrepAn interactive code search service focused on very fast indexed queries.pick this instead when search speed and a narrow operational footprint outweigh code chat and account features.
GitHub MCP ServerGitHub's MCP server for giving agents controlled access to repositories and platform operations.pick this instead when all code already lives on GitHub and agent access matters more than a separate multi-host search index.

What people are saying

  1. [github-trending] sourcebot-dev/sourcebot

Sources

  1. Sourcebot README
  2. Sourcebot v5.1.6 release
  3. Sourcebot license
  4. Sourcebot sizing guide
  5. Sourcebot roles and permissions guide
  6. Sourcebot MCP server guide
  7. Silent search limits issue
  8. Horizontal scaling RFC