Reference
Changelog
Release notes for webctx.
This changelog tracks code and product changes in webctx. It intentionally skips docs-site-only updates.
0.3.1 - 2026-11-13
- Added
WEBCTX_FIRECRAWL_PLANfor Free, Hobby, Standard, Growth, and Scale subscriptions. Both the CLI and Origo MCP now use the selected plan’s published limits for Firecrawl scrape, map, search, crawl, agent/extract, batch scrape, status, and browser-session operations. - Scaled local browser concurrency with plan capacity and replaced one-token-per-request Firecrawl rate files with compact, cross-process rolling-window counters that keep working at high paid-tier request rates.
- Preserved conservative Free limits when the setting is absent or unrecognized, plus the existing per-account scheduling, explicit upstream 429 cooldowns, and paid-operation idempotency.
0.3.0 - 2026-10-30
- Added capacity-aware search scheduling across Tavily, Exa, and Brave. Each configured key uses its own request budget and cooldown, coordinated across concurrent WebCTX processes on the same machine. A throttled account can fail over to another healthy account.
- Added per-key rate-limit defaults, configurable admission waits, provider retry guidance, and adaptive Brave limits learned from response headers. Busy providers can return partial results without holding up all federated search.
- Added multiple Firecrawl accounts through
FIRECRAWL_API_KEYSand macOS Keychain management. Scrape, map, crawl, and browser resources are tracked per account, with failover to available accounts for supported operations. - Kept Firecrawl browser sessions and paid Alexandria executions on their originating account, protecting session behavior and paid-request idempotency.
- Added multi-process quota tests, account-failover tests, and shared browser-capacity regression coverage.
0.2.11 - 2026-10-29
- Added Windows x86 (32-bit) and ARM64 to the release build matrix, alongside existing Windows x64, macOS, and Linux targets.
- Updated the npm installer to select the native Windows x86 binary when Node.js reports
ia32, while keeping the existing x64 and ARM64 mappings. Added platform-resolution regression tests.
0.2.10 - 2026-10-29
- Added Windows x86 (32-bit) and ARM64 GitHub release binaries after initial publication, alongside the existing Windows x64 build. npm installation support for Windows x86 starts in v0.2.11.
- Added multiple API keys per federated search provider:
BRAVE_API_KEYS,TAVILY_API_KEYS, andEXA_API_KEYSaccept comma/newline-separated credentials or JSON arrays, alongside the backward-compatible singular key. Duplicates are removed; plural keys are preferred over legacy singular credentials. - Automatically tries the next configured key for Brave, Tavily, or Exa when the provider rejects a key with an authentication, credit/quota, rate-limit, or server error (or a transport error). Bad requests and malformed successful responses are not retried. Provider failures cannot leak key values in CLI error text.
- Recognizes Brave’s unusual HTTP 422
SUBSCRIPTION_TOKEN_INVALIDresponse as a key error, without retrying unrelated 422 request-validation failures. - Added
webctx keys add <provider> --stdin,keys remove <provider> <index>, andkeys listfor Keychain-backed managed credentials on macOS. Newest managed keys are preferred, and status reports key counts without exposing secrets. Non-macOS users can configure plural environment variables directly. - Kept the separately configured Firecrawl/Alexandria integration and paid execution semantics unchanged; these new key-rotation rules apply to Brave, Tavily, and Exa search only.
0.2.9 - 2026-10-29
- Added persistent
webctx providers enable|disable <brave|tavily|exa>controls and awebctx statusoverview of enabled search providers and locally configured credentials, without exposing API keys or making live quota checks. Disabled providers are never queried even if their keys remain installed. - Kept normal search working with any enabled provider. Keyword searches prefer Exa’s include-text search, but now fall back to a clearly labeled Brave/Tavily quoted-phrase query when Exa is disabled, unconfigured, or fails (not equivalent to a strict page-text filter).
- Fixed Brave Search JSON decoding by leaving gzip negotiation to Go’s HTTP transport; a manually set gzip header previously caused compressed responses to be parsed as raw JSON.
- Added regression tests for persistence, safe error handling, disabled-provider request suppression, single-provider search, and Exa credit-exhaustion fallback.
0.2.8 - 2026-10-12
- Serialized identical paid requests within each warm Origo worker, preventing concurrent same-ID executions from racing before a replay receipt is stored. Partial batches are cached as well: replaying the same request no longer resubmits its successful charged operations. Cached replays report zero new provider executions; cross-worker billing remains explicitly unverified.
- Repaired filtered-source discovery pagination beyond 100 operations. Each page is ranked locally and the result explicitly identifies its ranking scope and whether additional source operations remain unvisited. Unfiltered semantic search now rejects a ranked offset that cannot be served by the provider, rather than silently changing discovery modes.
- Strengthened per-record credit preflight for operations exposing multiple result-count fields by using the largest explicit bound, not the first. Supported single-record FullEnrich mobile enrichment according to its inspected one-record response contract while keeping all provider-specific required-identifier validation.
- Explicit
map_site.limit=0(and CLI--limit 0) now returnsinvalid_limitinstead of silently using the default, matching the documented 1–100 range. - Added live-contract and automated regression coverage for cross-page filtered discovery, simultaneous identical paid requests, partial-batch replay safety, high-credit singleton pricing, and explicit zero limits. Preserved the original five MCP tools, server-controlled 200-credit maximum, and absence of model-facing budgets.
0.2.7 - 2026-10-10
- Audited the complete Origo MCP data workflow against the extended v0.2.6 quality report. Repaired filtered discovery (exact ID, source-domain lookup, transparent filtered ranking, task-oriented reranking, strict explicit zero limits) and made catalogue per-operation continuations consistently point to
inspect. - Improved execution safety: cost bounds recognize
k, page sizes, array counts and single-record lookups; operations above the server’s configured maximum are marked unavailable; multi-provider calls execute independently so one provider’s terms rejection no longer aborts unrelated data. - Added machine-readable rate-limit reset times, per-worker read throttling and free catalogue caching. Upstream quotas across independent serverless workers still depend on Firecrawl’s account limits.
- Added provider-specific pagination continuations for Treasury next-page links, ClinicalTrials.gov page tokens, Semantic Scholar continuation cursors, and FullEnrich
search_after. Reuses exact original query filters. - Made generated examples more conservative: clearly indicate alternative required inputs, distinguish structural validation from practical executability, avoid fake high-cost profile URLs, and include normalized integer, dependency and nullable-output contract hints without losing native schema fields.
- Clarified batch and operation receipts, different credit units, partial-result accounting and idempotent retries. Repeat requests handled by the same warm worker can be served from an in-memory replay cache with zero new upstream calls; cold-worker replay billing is explicitly unverified. Request-ID collisions no longer suggest repeating the conflicting payload.
- Replaced massive
map_sitedumps with a bounded 30-page default, filters for topic, path prefix and language, structured metadata and paginatednextcalls. Existing URL-only CLI mapping remains available but bounded. - Improved focused GitHub license questions with concise source-grounded answers and returned structured
invalid_urlerrors for unsupported URL schemes. Added regression tests for each issue family.
0.2.6 - 2026-10-10
- Removed user-controlled credit budgets and the paid-mode toggle from Webctx/Origo execution. The server enforces its own 200-credit per-request preflight ceiling, configured in Origo’s Vercel production environment, while returning both estimated and provider-reported actual credits.
- Completed the v0.2.5 ChatGPT MCP test-report improvements: sort discovery by reported similarity, generate clearly labelled schema-valid illustrative examples when a provider offers none, use typed discovery enums and non-null bounded call arrays, and return actionable validation errors as ordinary structured MCP responses.
- Made focused API documentation responses shorter and placed source-grounded REST contracts first; repaired copy-breaking line-continuation slashes inside JSON cURL payloads only when the corrected payload parses as valid JSON, and clearly annotated the correction.
- Added CPI observation measurement notes distinguishing index levels from year-over-year inflation percentages. Preserved exact provider observations, dates and provenance.
- Expanded regression coverage for server-enforced credit limits, absence of client budgets, generated examples, relevance, tool schemas, error envelopes and copyable REST examples.
0.2.5 - 2026-10-10
- Added three agent-friendly structured-data commands to both the Webctx CLI and Origo MCP:
research(semantic source and operation discovery, filters, browse and pagination),inspect(exact operation inputs, output contract, allowed values, examples and pricing), andexecute(paid provider-native data queries). - All three tools share the same Go engine and retain Firecrawl Alexandria’s full provider/category/group filters, exact operation IDs, native continuation requests, response schemas, batch execution of up to ten operations, and raw upstream JSON access.
- Tool results provide the exact next tool name and argument shape. Contract inspection supplies input templates, required field hints and the next execute call; validation errors describe how to fix missing, invalid or out-of-range inputs before spending credits.
- Paid execution uses one request ID for idempotent retries, a configurable per-request credit preflight cap, provider-native provenance and pagination continuations. Third-party terms cannot be accepted automatically; Firecrawl team limits remain authoritative.
- Preserved the existing
read_link/map_siteMCP contracts and the three original CLI commands. The hosted Origo interface now contains five MCP tools, with chargedexecuteclearly distinguished from free discovery and reads.
0.2.4 - 2026-10-10
- Fixed native Markdown retrieval for documentation sites: use bounded GET requests rather than positive HEAD content-length assumptions, and recognize directory index.md routes. Canonical Stripe and Cloudflare URLs now prefer publisher Markdown while retaining the original URL and precise Markdown source URL.
- Hardened long CAPTCHA/challenge detection before accepting Firecrawl or browser fallback responses. Failed challenge resolution is reported instead of leaking hCaptcha boilerplate; optional WEBCTX_DEBUG=1 traces retrieval stage, outcome, and duration without URL query credentials.
- Improved question-driven exploration with publisher llms.txt indices, same-topic prioritization, getting-started guides, REST documentation, webhook instructions, and follow-ups when explicit requested evidence is missing.
- Preserved complete code fences and fixed excerpt-budget escape for JSON-rich Markdown. Legacy migration content is demoted for new-project questions, and missing requested fields are called out rather than implied.
- Added regression tests based on live Origo MCP cases covering Stripe Checkout, Cloudflare Durable Objects, and Alexandria REST documentation. Origo continues exposing exactly two read-only tools and sharing the Webctx Go backend.
0.2.3 - 2026-10-10
- Improved question-focused reads in Webctx CLI and Origo MCP: complete code, cURL, and API examples are retained rather than truncated in the middle of a fence.
- Follow the published llms.txt index to the canonical OpenAPI specification for REST/API questions. Include exact method and path contracts, schema fields, and published request payloads, with authoritative source URLs.
- Fixed the Alexandria case: discovery and execution examples stay intact, and the missing firecrawl/find-tools inspection REST example comes from Firecrawl’s published OpenAPI specification.
- Preserve existing URL-only behavior and Origo’s two-tool MCP interface; no paid Alexandria provider execution is performed.
0.2.2 - 2026-10-09
- Added the Origo MCP transport to the shared Go retrieval engine, exposing
read_linkandmap_sitewithout changing the CLI’ssearch,read-link, ormap-sitecommands. - Improved Firecrawl read-link recovery with staged proxy escalation and a disposable browser fallback for inaccessible or blocked pages.
- Enabled Firecrawl’s built-in 30-minute scrape cache (
maxAge) without adding an application-managed cache or database. - Hardened URL validation against local/private addresses, rejected empty or challenge-page extractions, and stopped bypassing TLS certificate verification during Firecrawl scrapes.
- Fixed the browser fallback to use supported DOM text extraction and added MCP, proxy, browser, and URL-validation regression tests.
0.2.1 - 2026-08-15
- Completed native GitHub context coverage across issues, pull requests, commits, Actions, releases, Discussions, Gists, packages, projects, search, profiles, activity, and deployments.
- Bounded large GitHub roots and provider fan-out so broad reads remain useful without overwhelming the default context.
- Closed the GitHub read-link review gaps and documented the cross-family acceptance coverage.
0.2.0 — 2026-08-14
- Added native GitHub reads for repositories, issues, pull requests, commits, Actions, releases, Discussions, Gists, packages, projects, search, profiles, activity, and deployments.
- Added focused GitHub subresource views, bounded pagination, release assets, job logs and artifacts, history, compare, and blame support.
- Made large GitHub roots navigation-first across Issues, PRs, commits, releases, trees, Discussions, Gists, statistics, deployments, lists, Search, Packages, Projects, and workflow/history views, with exact copied selectors and explicit raw
.diff/.patchdrill-downs. - Separated GitHub/provider incompleteness from local context omission, and bounded long human-authored index metadata so large provider pages cannot amplify a default read.
- Reconciled anonymous and authenticated GitHub behavior while preserving crawler fallback for public package pages.
- Preserved insertion order when search results receive tied scores.
0.1.1 — 2026-03-22
- Fixed credential loading from binary environment variables and the OS keychain.
- Ignored the local
tmpworkspace.
0.1.0 — 2026-03-22
- Ported the webctx CLI to Go.
- Prepared the npm release flow for distributing the Go binary through npm.
- Fixed build-all artifact names.
- Added an environment example for provider credentials.