We publish what shipped, not what we hoped would ship. If a feature rolled back, you'll see the roll-back. If a number regressed, you'll see the regression. Builders read changelogs to know what they can trust — this one respects that.
The six built-in presets are no longer the ceiling. Send a custom state machine inline on POST /v1/agents — your own states, transitions, allowed_tools per state, parallel_group fan-out, all the same fields the built-ins use. Server enforces hard caps: max 50 states, ≤256 KB JSON, no transition cycles, every allowed_tool must be a tool the runtime actually exposes. Cycle detection is a real DFS, not a comment.
Four conventional namespaces — short_term, long_term, entity, contextual — now have first-class helpers in the Python, TypeScript and Rust SDKs. Same vocabulary as LangChain and CrewAI. The generic /v1/memory/:namespace endpoint is unchanged: pick the convention or your own name.
The live agent sandbox now renders a Gantt-style timeline of the state machine — each step's duration is a bar, success is green, retries amber, cache hits blue, failures red. Hover gives token counts and the next state. Open the capot during a pitch and the orchestrator stops being a black box.
The live agent sandbox at /tools/test-agent is back — pick a preset, write a prompt, see a real verdict and per-claim breakdown without signup. The docs page now spells out the wauldo.toml schema and the available presets, so a new dev can configure an agent in one read instead of guessing.
Deep audit pass over the agent runtime: verdict and claim counts now reflect the real fact-check (no more cosmetic numbers), tighter SSRF protection on every web tool, broken MCP connections surface a clear error, background task crashes log instead of disappearing.
Connect any MCP server and every remote tool becomes a Wauldo tool, registered alongside the natives. The whole MCP ecosystem is now reachable through one config file.
Four more native tools. Read feeds, post to Slack, read GitHub repos and issues, send email via SMTP. Closing the gap with the bigger frameworks, ship by ship.
Three more native tools for agents. Encyclopedic lookups, academic paper search, and date arithmetic.
Agents can now search the live web. One more tool, fewer reasons to bring your own framework.
The RapidAPI path now returns a per-claim breakdown via a base64-encoded JSON header. VSCode plugin users on the marketplace path get the full verification drawer with claims and source cards — the same view Enterprise direct users already had.
Verdict mapping reworked. CONFLICT is now reserved for actual contradictions (hallucination rate above 0.5), not low support scores. UNVERIFIED means "no source matched" — no more fake PARTIAL verdicts when the lexical layer couldn't find a hit.
Dual-mode client that auto-detects RapidAPI marketplace URLs versus direct Enterprise endpoints. Inline verify drawer surfaces claims and sources next to any answer. Packaged locally at v0.5.0, ready for marketplace when the founder says go.
Python SDK shipped async support end-to-end. Seven community pull requests merged this cycle. Thanks to @qorexdev (7 PRs) and @dagangtj (1 PR) for the outside contributions.
The /v1/upload endpoint now accepts PDF and DOCX files directly. Automatic chunking, multilingual text extraction across French, English, and Spanish. No preprocessing required on the client side.
Adversarial pass rate jumped from 81% to 97% after we shipped verification inside the reasoning loop instead of bolting it on after the fact. The 0% hallucination rate held across the regression suite.
Python landed on PyPI, TypeScript on npm, Rust on crates.io. All three MIT-licensed. Identical surface across languages — write once, port the client in minutes.
wauldo.com live. api.wauldo.com in production. Session state and chunk storage cached behind the API. First external users onboarded same day.
Seven LLM providers behind one router. Cost-aware fallback between them based on latency and price per token. Tiered intelligence selection per query — cheap for trivial lookups, strong for reasoning chains.
Typed tenant identifiers enforced at every request boundary — no raw strings. API key revocation list layered on top of the issuance store. Auth failures default to closed, never open. Three-layer defense against prompt injection attacks on ingested content.
Initial backend. First version of POST /v1/fact-check. Basic retrieval-augmented generation with a hybrid lexical-plus-vector scorer. The foundation everything else sits on.
New releases ship weekly or better. Subscribe via RSS to get every changelog entry in your reader. No email capture, no newsletter.
Subscribe to RSS →Six adapters, seventy adversarial tests, weighted Wilson intervals. Live results.
View leaderboard →Eval at 77%, hard suite at 85%, RAG adversarial at 100%. Reproducible runs.
Read benchmarks →Write-ups on routing, verification loops, hallucination rates, and retrieval.
Read the blog →Read the full git history on GitHub if you want the fine grain.