Changelog

1.0.0-preview.29

August 17, 2026 · Current preview

Preview 29 makes Scout more selective about what it returns and more conservative about what it claims: incompatible indexes cannot serve evidence, compact context suppresses exact clones, and uncertain relationships stop before they become invented topology.

• Indexes built by an incompatible extractor generation now fail closed. symbol_schema remains available for diagnosis while every evidence-serving tool returns a machine-readable refresh-required error until reconciliation is healthy.
• Compact fast_context now prioritizes direct implementation evidence, discovers likely tests from the strongest production files, hydrates read ranges around the best match, and collapses exact header, vendor, fixture, and tool clones to one canonical representative. Expanded detail and exact symbol tools still expose every indexed path.
• Graph, trace, path, and impact traversal now admit only receiver-typed or proven language-rule edges. Name-only and unknown relationships remain visible at the first hop without manufacturing downstream certainty.
• Heuristic relationship targets are revoked and recomputed after relevant index changes, receiver registries stay inside one language family, and an unresolved typed receiver cannot fall through to an unrelated globally unique function.
• C and C++ extraction recovers declarations after kernel-style loop macros without overstating coverage. Persisted call argument counts reject only provably incompatible fixed-arity overloads; defaults, variadics, operators, and uncertain signatures remain unresolved.
• Python and PHP runtime-name guards preserve negative knowledge when the likely implementation is outside the workspace, while import evidence and deliberate same-file shadows still win. Precision blockers are explained in reference results.
• The Claude integration and bundled agent guidance now start broad work with one compact fast_context call, read cited ranges before more navigation, omit unnecessary initial detail and limit expansion, and stop when the evidence is sufficient.
• The 120-run real-world campaign found a small correctness advantage for Scout and a clear structural-navigation win on AT002, but also exposed expensive correct solutions on AT005. Preview 29 directly addresses the observed stale-index, clone-diversity, and downstream-certainty mechanisms; generalization claims remain gated on separately frozen unseen tasks.
• All four native candidate lanes completed 514 tests with zero failures and six intentional live/archive ignores on macOS and Linux; Windows completed 483 with zero failures and five ignores. The exact signed release was privately certified, promoted, and independently downloaded and verified while the stable channel remained untouched.
• Signed Preview 29 wrappers for Codex and Claude Code each passed a clean live-agent symbol search. Four target-specific MCPBs are active/latest in the Official MCP Registry, Smithery serves the byte-identical universal bundle with all fourteen tools, and the existing MCP.so submission now points at Preview 29. The Preview 29 skills-only package is published in the OpenAI Plugins Directory; Anthropic’s directory review remains explicitly pending.

Read the Preview 29 release report →

1.0.0-preview.28

August 8, 2026 · Previous public preview

Preview 28 turns independent agent feedback into a tighter tool contract: question-form context queries work, scoped schemas validate, compact responses cost fewer tokens, and ranking distinguishes production code from test and fixture noise.

• fast_context now converts ordinary question phrasing into deterministic local keyword bundles, so a question such as how the MCP server authorizes a tool call reaches the authorization chain instead of returning an empty shortlist.
• Scoped symbol_schema gap entries always include diagnostics, including an empty array, so runtime results conform to the advertised output schema and directory health checks no longer fail validation.
• Default symbol records now contain only stable locator fields; references omit redundant symbol hydration, and architecture responses normalize repeated paths into compact tuple rows. Expanded detail remains available when diagnostic metadata matters.
• symbol_references checks all indexed relationship kinds by default, including uses_type, and empty answers state what was checked. Rust struct-field type uses now contribute relationships and retain exact import bindings through Cargo dependency identity.
• Search and architecture ranking share a production/test/fixture/example/generated/script classifier. Production symbols win ties without hiding explicitly scoped tests, and inline Rust test modules are classified as test hotspots even inside src files.
• Outline results report exact pre-limit match counts and source ordering, with concrete kinds, nesting, and limit recovery guidance. Scoped schema responses include only the language capabilities relevant to that scope.
• All 501 release tests passed with six intentional live/archive ignores and zero failures. Four native packages and all immutable public release objects passed certification, promotion, and independent verification; the stable channel remains untouched.
• The signed Claude and Codex wrappers were published, with Codex validated through installation and a live navigation check. Four target-specific MCPBs were independently verified on GitHub, Preview 28 became active/latest in the Official MCP Registry, and Smithery published the exact unchanged universal bundle with all fourteen public schemas verified. The existing MCP.so submission now points at Preview 28 and remains under free review.

Read the Preview 28 release report →

1.0.0-preview.27

August 4, 2026 · Previous public preview

Preview 27 makes exclusions visible instead of silent, teaches the Claude Code hooks to verify an index exists before claiming one does, and gives the integration one command that keeps the installed plugin, marketplace registration, and managed memory aligned with the runtime.

• A workspace .atlasignore file provides Scout-only exclusions that do not affect other ignore-aware tools, and .gitignore files are honored even outside a Git repository, matching ripgrep’s --no-require-git behavior. Editing a known ignore file now forces the ignore-aware discovery pass on the next reconcile instead of waiting for an unrelated full walk.
• Exclusions are reported instead of appearing as unexplained coverage gaps. Discovery records excluded paths at their shallowest roots, capped at 256 entries with an explicit truncation flag, and symbol_schema returns the list as an optional ignored_paths block filtered to a requested directory scope.
• The Claude Code hooks stop claiming a workspace is indexed without looking: both the tool notice and the new prompt reminder verify that an index database exists at or above the hook’s working directory and stay silent when it does not.
• The search notice now answers the searches sessions actually run. It parses Bash commands for rg and grep-family invocations in addition to the Grep tool, applies the same bare-identifier gate, and raises its per-session budget from two to four. A new UserPromptSubmit hook repeats the navigation policy on the first prompt and every fifth after, only in indexed workspaces.
• atlas-scout claude-code sync aligns the managed CLAUDE.md memory section, the marketplace registration, and the installed plugin with the current runtime, reporting each independently and never touching content outside its markers. atlas-scout upgrade runs the upgraded binary’s sync after promotion, and doctor recommendations point at the command.
• The SessionStart hook spawns a detached index refresh for the session’s working directory, so the first structural query answers from a current index without the session waiting. Directories without a .git or .atlas ancestor are left alone.
• A reconcile runs a best-effort VACUUM when the SQLite freelist exceeds both 16 MiB and one quarter of the database, returning pages to the filesystem after a workspace shrinks - without letting optional space reclamation become an indexing failure.

Read the Preview 27 release report →

1.0.0-preview.26

August 3, 2026 · Previous public preview

Preview 26 gives modern MCP clients substantially more useful evidence in a smaller response, removes broad classes of false relationships, and makes indexed Go reconciliation faster without hiding the costs or uncertainty behind those gains.

• Modern MCP clients now receive the canonical structured result without a byte-identical JSON text mirror. A measured fast-context response carried 71% more evidence in an 18% smaller envelope; legacy clients retain the mirror without losing evidence to the budget.
• Relationship resolution now stays inside language-family namespaces, preventing Rust, TypeScript, Python, and fixture declarations with the same name from binding to one another. Every resolved edge reports a named evidence tier instead of asking callers to interpret a decimal confidence value.
• Unconfirmed method-name matches remain visible but no longer extend graphs, traces, paths, or multi-hop impact. Receiver typing recovered the relationships it could prove, while common standard-library method names remain unresolved rather than acquiring false callers.
• Go package clauses, imports, and build constraints are recorded during extraction instead of rebuilding context by parsing the complete Go corpus again. On a 996-file Go workspace, resolution fell 41% and full indexing fell 15%, with the old and new context verified field by field.
• Scoped queries now use SQLite index range seeks, reads can continue through a write transaction, prepared file cleanup became 5.1× faster on the audit workspace, overlay-aware pages stay full, and truncation and hotspot completeness are reported truthfully.
• Claude Code can now receive a bounded advisory hint when it is about to grep for a code identifier, and fast context plus edit impact join the eagerly visible tool entry points. The hook never makes a permission decision and ignores prose, regex, and non-Grep calls.
• All 474 release tests passed with six intentional live/archive ignores and zero failures. Four native packages and all 24 immutable public objects passed certification, promotion, and independent verification; the stable channel remains untouched.

Read the Preview 26 optimization report →

1.0.0-preview.25

August 1, 2026 · Previous public preview

Preview 25 prepares Atlas Scout’s Claude Code integration for Anthropic’s community plugin marketplace and makes its guidance more precise about when Scout helps, where coverage is partial, and who chooses the right tool.

• The Claude Code integration now has a public, inspectable Apache-2.0 wrapper with validated marketplace metadata, installation guidance, and security documentation. The proprietary Atlas Scout runtime remains a separate local installation.
• The community-directory submission is under Anthropic review. Until it is published there, Claude Code users can install the wrapper directly from the Zaguán Labs GitHub marketplace source or continue using the verified local marketplace installed with Scout.
• Plugin guidance now describes Scout as an optional local navigation aid rather than a universal replacement for Claude’s built-in tools. It preserves language-coverage caveats and leaves tool choice to Claude and the user’s request.
• Session and subagent hooks remain advisory only: they do not inspect conversation history, execute repository code, approve or deny tools, modify calls, or send source code to Zaguán Labs.
• All four native packages passed the release pipeline, and all 24 immutable public objects were independently downloaded and verified. The stable channel remains untouched.

Read the Claude Code plugin guide →

1.0.0-preview.24

August 1, 2026 · Previous public preview

Preview 24 removes most of the serialized SQLite work from fresh indexing, cutting complete readiness by more than half on locked Linux and Firefox campaigns without changing the logical index or search results.

• Scout now prepares validated database-ready records before they enter the single SQLite writer, allowing safe preparation work to overlap extraction instead of extending the serialized path.
• Relationship sources bind to symbol rows already established in the same atomic write, while relationships, symbols, and identifier terms use measured multi-row statements bounded by SQLite’s active runtime limits.
• The final 256-row design reduced median SQLite writer time by 67.13% on Linux and 60.86% on Firefox. Verified complete readiness improved by 57.72% and 54.44%, respectively.
• Canonical logical tables, ordinary and identifier full-text-search results, normalized MCP responses, transaction rollback, concurrent-reader behavior, no-change reconciliation, and one-file incremental updates remained inside their exactness and latency gates.
• The faster path uses more CPU while active. Peak memory rose 1.51% on Linux and 0.51% on Firefox, and one Firefox device-time counter regressed even as bytes, operations, flushes, and end-to-end duration fell.
• All four native packages passed the release pipeline, and all 24 immutable public objects were independently downloaded and verified. The stable channel remains untouched.

Read the Preview 24 performance report →

1.0.0-preview.23

July 31, 2026 · Previous public preview

Preview 23 gives HTML and PHP first-class structural extraction, carries that richer markup awareness into modern web frameworks, and makes Claude Code plugin upgrades easier to trust and diagnose.

• HTML is now parsed structurally rather than scanned as bounded text. Scout can navigate document IDs, custom elements, script and stylesheet dependencies, and HTMX request targets.
• PHP now contributes namespaces, imports, classes, interfaces, traits, enums, members, signatures, calls, type and inheritance relationships, trait use, Laravel-style routes, and surrounding template markup.
• Astro combines its TypeScript frontmatter with HTML markup extraction. Vue and Svelte templates use the same HTML path, while JSX and TSX now recognize custom elements and HTMX targets as well.
• The bundled Claude Code plugin now carries the Atlas Scout product version, preventing an older cached plugin from silently retaining stale hooks or MCP configuration after an upgrade.
• atlas-scout doctor now reports the preferred plugin state, its marketplace source, stale installed versions, and competing manual MCP registrations so integration problems are visible instead of ambiguous.
• All four native packages passed the release pipeline, and all 24 immutable public objects were independently downloaded and verified. The stable channel remains untouched.

Explore the expanded web-language support →

1.0.0-preview.22

July 30, 2026 · Previous public preview

Preview 22 makes Atlas Scout materially lighter on large repositories while preserving the complete structural map, warm-query performance, and verified readiness.

• Fresh schema-6 indexes store relationship sources through compact internal row IDs while preserving the stable symbol identities, source paths, target identities, and latency-critical indexes exposed to agents.
• The indexing pipeline no longer allows a fully expanded batch to wait in memory while another is being extracted. Extraction and database writing still overlap, but large C and C++ relationship batches no longer accumulate in the handoff.
• Across three formal fresh runs each on Linux and Firefox, final cache size fell 25.64–30.28%, peak cgroup memory fell 21.85–24.75%, and physical writes fell 24.79–27.82%. Verified complete readiness also improved 3.01–3.10%.
• All six formal positions completed, passed their asserted MCP readiness checks, used zero swap, and retained identical indexed content. The measured 230-request warm-query stream remained 0.49 seconds p50 before and after.
• Existing indexes migrate atomically without a readiness-blocking vacuum. Fresh indexes receive the smaller physical representation immediately; upgraded indexes keep freed pages available for reuse.
• Scout now adopts MCP 2026-07-28 with stateless discovery, per-request protocol-version metadata, modern result and cache metadata, and explicit unsupported-version errors while preserving legacy initialization and response shapes.

Read the Preview 22 release notes →

1.0.0-preview.21

July 29, 2026 · Previous public preview

Preview 21 keeps MCP initialization reachable when Antigravity 1.1.8 sends a roots-changed notification before the standard initialization request.

• Scout now ignores an early notifications/roots/list_changed message until the client sends the required initialize request, instead of letting the premature notification disrupt startup.
• After initialization, the ordinary MCP session continues on the same connection; users do not need a special Antigravity launch command or transport configuration.
• The existing server/discover compatibility fallback remains intact: unsupported discovery probes receive the standard method-not-found response without closing the connection.
• All four native packages passed tests, release builds, offline smoke tests, SBOM generation, Sigstore attestation, OpenPGP certification, private staging, production promotion, and public verification.
• All 24 immutable public objects were independently checked against the release record. The stable channel remains untouched.

Read the MCP host guide →

1.0.0-preview.20

July 28, 2026 · Previous public preview

Preview 20 is a focused Go correctness hot-fix that keeps same-named methods on different receiver types distinct throughout indexing and caller navigation.

• Concrete Go method identities now include the normalized receiver type, so methods with the same name on different types receive different stable symbol IDs.
• The corrected identity prevents one receiver method from replacing another definition, preserves their separate caller identities, and stops typed calls from hydrating as the wrong method.
• The Go extractor advances so existing indexes re-extract unchanged Go files once after upgrade and rebuild the corrected method identities automatically.
• The release also includes the race-safe Windows transcript expectation prepared after Preview 19. It accepts documented transient health metadata during background revalidation without changing runtime output.
• All four native packages passed the complete release pipeline, and all 24 immutable public objects, checksums, cache policies, and OpenPGP signatures were independently verified. The stable channel remains untouched.

Read the Go support boundaries →

1.0.0-preview.19

July 28, 2026 · Previous public preview

Preview 19 keeps normal startup responsive on large, shared indexes without weakening explicit database verification or the health evidence agents rely on.

• Normal CLI, MCP, and daemon startup no longer waits for a full physical database scan or a complete recount of the index before it can proceed.
• After a successful reconciliation, Scout saves one atomic healthy count snapshot and restores it on later opens. Startup health remains informative without repeatedly scanning the complete index.
• The explicit atlas-scout index maintenance path still performs synchronous physical verification and quarantines a corrupt generated database before rebuilding it.
• Several agents can now open the same large workspace index without serial integrity and count scans delaying or timing out concurrent MCP initialization.
• Regression coverage exercises normal and verified opens, restored counts, simultaneous core opens, independent MCP processes, and preserved corruption detection.
• All four native platforms passed, and all 24 immutable public artifacts were independently checked for signatures, hashes, sizes, and cache behavior. The stable channel remains untouched.
• A post-tag source correction broadens one Windows test expectation to accept documented transient health states while background validation finishes. The released Preview 19 runtime was already correct and its bytes did not change.

Read the startup and recovery guide →

1.0.0-preview.18

July 27, 2026 · Previous public preview

Preview 18 substantially reduces the MCP tool catalog while preserving exact runtime results and improving aggregate agent validity in a frozen 224-capture study.

• The complete compact catalog falls 58.59% for Pro and 56.41% for Free without removing a tool, changing edition entitlements, or weakening any tool-specific data result.
• Repeated catalog prose becomes stable edition-aware instructions: Free guidance names only Free tools, while both editions retain the rules that protect exact IDs, bounded reads, partial coverage, raw fallback, and response language.
• The advertised metadata schema becomes smaller and forward-compatible while retaining all six metadata fields and the complete empty-result trust vocabulary. Runtime structured results and text fallbacks remain unchanged.
• Three evidence-driven input guardrails state the four-query fast-context limit, clarify that outlines are not cursor-paginated, and put relationship kinds on reference queries. They cost 1,067 bytes and were kept because smaller-model routing improved.
• Across 112 control and 112 treatment captures, strict validity rises from 103 to 107. Codex, Claude Code, and Antigravity are 70/70 valid; aggregate OpenCode validity improves from 35/42 to 37/42.
• Every treatment capture produces a substantive final answer, every measured host/model aggregate uses fewer Scout calls, and no run times out, leaks compressed schema grammar, or answers in the wrong language.
• The signed four-platform release and all 24 immutable public objects passed independent signature, hash, size, and cache-policy verification. The stable channel remains untouched.
• The report retains adverse rows: Kimi and DeepSeek are slightly slower in aggregate, DeepSeek uses 1.60% more input, and several individual small-model tasks regress materially.
• Daemon protocol 4 prevents an upgraded client from silently receiving the older catalog through a stale same-major daemon; direct and daemon catalogs remain byte-equivalent.

Read the 224-capture benchmark →

1.0.0-preview.17

July 26, 2026 · Previous public preview

Preview 17 gives Atlas Scout a native, signed update path that keeps its runtime and integrations on one verified release.

• A new atlas-scout upgrade --check command verifies the signed preview channel and reports whether an update is available without downloading an archive or changing the installation.
• atlas-scout upgrade performs the separate, user-approved installation step. Agent and automation workflows can use strict JSON output and must accept the unsigned preview explicitly before changing anything.
• On Linux and macOS, the updater verifies release identity, target, signatures, archive length, and SHA-256 before staging the complete platform release and smoke-testing both executables.
• The runtime pair, Claude Code plugin, documentation, and legal notices move together through one atomic current-release selector. Existing indexes, licenses, configuration, and workspace files are left untouched.
• Validated older flat installations can migrate into the managed layout while retaining their prior executables. Unmanaged, package-manager, conflicting, or unsafe installations are refused instead of being modified by guesswork.
• Running daemons and MCP sessions keep using their previous executable until restarted, so the updater reports the required restart rather than terminating active work.
• All four native platforms passed tests, release builds, offline smoke tests, packaging, SBOM generation, provenance, OpenPGP certification, private staging, and public verification.
• Preview 15 and Preview 16 remain immutable rejected candidates. Preview 17 carries their intended behavior after the macOS path and Windows signed-fixture defects were corrected and validated.

Read the upgrade and lifecycle guide →

1.0.0-preview.14

July 25, 2026 · Previous public preview

Preview 14 teaches Scout to follow a conservative subset of Go and C++ calls while making routine MCP responses substantially quieter.

• Go calls can resolve through declared packages, exact indexed imports, test-package boundaries, conservative build constraints, and simple statically proven receiver types.
• C and C++ indexing now retains literal includes and structured call observations; a precision-first C++ subset resolves through lexical ownership, simple receiver types, same-file definitions, and direct include visibility.
• Go and C++ calls no longer fall back to an unrelated language or a project-wide same-name symbol when their own evidence is insufficient.
• Healthy, complete MCP results now omit routine diagnostic fields. Empty, partial, stale, degraded, repaired, or truncated results still carry the evidence an agent needs to judge them.
• The Pro tool catalog is 45.2% smaller, the Free catalog 43.9% smaller, and healthy exact-result metadata falls from 635 bytes to an empty two-byte object.
• Search guidance now favors the compact five-result default and requests connected evidence only when the preview is likely to replace another structural call.
• Go and C/C++ files re-extract once after upgrade so the new semantic observations are available. The local daemon protocol advances to 3 to keep older response contracts from being served silently.
• The embedded agent workflow now stays within its fixed wire budget after Windows line-ending conversion, closing the cross-platform acceptance gap that rejected the immutable Preview 13 candidate.
• Preview 14 passed 376 workspace tests, four agent-benchmark analyzer tests, the VS Code compile, and version synchronization checks; five live-service tests remain intentionally ignored.

Read the Preview 14 evidence →

1.0.0-preview.12

July 24, 2026 · Previous public preview

Preview 12 gives Rust agents precise qualified-call relationships and prevents architecture summaries from turning bounded requests into runaway database work.

• Qualified Rust calls through Self, crate and module paths, explicit imports and aliases, local re-exports, renamed workspace dependencies, and exact UFCS syntax can now resolve when Cargo and source evidence prove one target.
• Failed qualifiers, glob-derived visibility, missing dependencies, generated or macro-expanded code, and unproven trait selection remain unresolved instead of weakening into a same-name method guess.
• References and graph evidence preserve the qualified spelling and exact byte position, including two distinct calls written on the same source line.
• Reference results carry at most ten complete observations per page and provide a continuation cursor when more callers remain.
• Architecture summaries now aggregate file counts safely, honor cancellation, and stop after a bounded internal work period instead of continuing abandoned relational work.
• The generated database advances to schema 5. Existing Rust files are re-extracted once after upgrade; source files are never changed.
• Published evidence records perfect precision and recall over the labeled supported subset, faster Firefox and Linux indexing, and an openly disclosed 166 ms median regression on the small Tokio corpus.
• Preview 12 carries the same product behavior as the unpromoted Preview 11 candidate; only its cross-platform release acceptance check changed.

Read the Rust and architecture evidence →

1.0.0-preview.10

July 23, 2026 · Previous public preview

Preview 10 restores reliable startup with modern MCP clients that check for a newer discovery flow before beginning the supported initialization exchange.

• Clients such as Antigravity CLI can probe for server discovery before sending the normal MCP initialization request without losing the Atlas Scout connection.
• Scout answers the unsupported probe with the standard method-not-found response, then keeps the same connection ready for ordinary initialization and tool use.
• The compatibility behavior is identical for direct atlas-scout mcp sessions and sessions using the local atlas-scoutd daemon.
• No host configuration change is required after upgrading and starting a new MCP session.
• This is a connection-startup compatibility repair, not support for stateless MCP, Streamable HTTP, or the server/discover method itself.

1.0.0-preview.9

July 22, 2026 · Previous public preview

Preview 9 makes workspace moves and copies recover automatically without risking reuse of an index that belongs to the old location.

• When a workspace arrives at a new canonical path with its .atlas directory, Scout now preserves the old generated cache and creates a clean index for the new location.
• MCP initialization can continue without asking the user to find and delete stale index files manually.
• Scout never rewrites or opens the old SQLite and write-ahead-log state under the new identity, avoiding conflict with a process that may still use the original workspace.
• Diagnostics report preserved relocation archives so they can be inspected and removed later as ordinary ignored generated data.
• Invalid identity data, unsafe filesystem shapes, failed preservation, and explicit external-cache collisions still stop with an honest error instead of attempting a risky recovery.
• The Preview 8 number belonged to a rejected, unpromoted candidate. Preview 9 carries the same runtime repair after its cross-platform acceptance fixtures were made portable and passed on all four targets.

1.0.0-preview.7

July 22, 2026 · Previous public preview

Preview 7 makes multi-million-symbol repositories dramatically faster to index and explore while preserving Scout’s conservative treatment of ambiguous code.

• Exact symbol searches now stay interactive on very large repositories instead of repeatedly scanning the entire symbol map.
• Path, language, and symbol-kind constraints are applied early, so common or overloaded names resolve to the requested file more reliably.
• Broad discovery retains useful CamelCase, snake_case, acronym, and qualified-name matches while doing substantially less work.
• JavaScript and TypeScript cross-file relationship indexing now narrows plausible targets before applying the same precision-first acceptance rules; ambiguous matches remain unresolved rather than guessed.
• The release adds regression coverage for scoped names, broad identifier discovery, and large-project relationship equivalence.
• Locked Firefox and Linux benchmarks confirm the large-repository gains, preserved semantic projections, and the indexing and storage costs that come with the faster retrieval path.

Read the Firefox and Linux benchmarks →

1.0.0-preview.6

July 21, 2026 · Previous preview

Preview 6 improves how Scout follows imported code, rendered components, likely tests, and shared workspace activity.

• JavaScript, TypeScript, JSX, and TSX imports now contribute stronger call and component-usage evidence, including aliases and common project-root path shortcuts.
• Reference queries can target an exact symbol within a known file and can find the components that render a JSX or TSX component.
• Edit-impact results now include a small, clearly non-exhaustive shortlist of tests that are likely to matter.
• Source coordinates are more consistent, and health reports now explain when another Scout process is updating the shared index while reads remain available.
• Unrelated functions that happen to share a name are less likely to appear as false relationships.

1.0.0-preview.5

July 20, 2026 · Claude Code integration and smaller responses

Preview 5 makes Atlas Scout a first-class Claude Code integration and reduces the amount of context needed for common navigation work.

• Release archives now include a Claude Code plugin that makes Scout available to the main session and delegated subagents.
• Optional workflow guidance and compact task recipes help agents choose the right structural tool without preventing normal reads, searches, Git, tests, or terminal work.
• Search, outline, and broad task-orientation results start with smaller defaults and can still be expanded when deeper evidence is needed.
• Search ranking better separates useful identifiers from incidental path matches.
• Empty reference results describe their coverage instead of implying that a symbol is definitely unused.

1.0.0-preview.4

July 19, 2026 · Better Go and React understanding

Preview 4 expands relationship evidence for Go and React while making host discovery and normal responses more dependable.

• Go interface implementations can be discovered when the evidence is unambiguous.
• Rendered JSX and TSX components participate in references and change-impact analysis instead of appearing unused.
• File outlines are concise by default, with deeper declarations available on request.
• Codex, Claude Code, and generic MCP hosts receive clearer guidance for finding Scout without losing access to ordinary repository tools.
• Index upgrades, persisted health, stale identifiers, and ambiguous relationships now fail or recover more honestly.

1.0.0-preview.3

July 17, 2026 · Clearer setup and diagnostics

Preview 3 delivers the setup, licensing, and diagnostic improvements prepared for the unpublished Preview 2 candidate.

• Indexing and diagnostics now use readable terminal output by default, with JSON and legacy machine-readable forms available for automation.
• Device activation gives clearer labels, safer credential-store checks, and more useful recovery instructions when local storage fails.
• Host setup distinguishes global registrations from project-pinned configurations and reminds users to reconnect after configuration changes.
• Generated database size and bounded indexing progress are visible in the CLI.

1.0.0-preview.1

July 16, 2026 · First public preview

Preview 1 established Atlas Scout as a local, multi-language code map with permanent Free navigation and optional deeper Pro analysis.

• Fourteen bounded, read-only MCP tools cover symbol discovery, references, structural paths, architecture, task orientation, and edit impact.
• Each workspace gets one shared, gitignored index that can be reused by agents, the CLI, and the optional local daemon.
• Free works without an account or network connection; evaluations and paid access use locally verified leases with non-destructive fallback to Free.
• Activation, offline status, refresh, deactivation, and device management are available from the CLI with native credential storage on supported platforms.
• Discovery became more relevant and more compact, with bounded source ranges, stable pagination, explicit health, and honest treatment of unresolved evidence.
• The release was validated against the complete Firefox checkout and shipped with signed checksums, provenance, and third-party notices.