Tracking the evolution of Sparsi—the tool for packaging repeatable multi-step tasks into stable, cost-efficient MCP servers. Each update focuses on expanding the deterministic operator library and refining the tools that help you build audit-ready AI workloads.
First public release of Sparsi. Establishes the core DAG workflow execution engine, the deterministic op library, and the typed AI op suite, together with MCP integration, retrieval and citations, AI-assisted repair, and pluggable credential routing — everything needed to package a business process as a deterministic, fully auditable tool, with AI confined to where judgment is genuinely required.
Core engine
dagor execution engine with topological scheduling, parallel branch
execution, skip-propagation for predicate gates, and vertex-level error attribution.
Deterministic op library
80+ typed ops covering arithmetic (float and integer variants — AddFloatOp /
AddIntOp, PowFloatOp / PowIntOp,
ModFloatOp / ModIntOp, clamp, sum, min/max), string
manipulation and casts (Float64ToStringOp, IntToStringOp,
BoolToStringOp, ToStringOp), slice operations, predicates,
boolean logic, JSON extraction, HTTP I/O, and select/coalesce patterns.
AI op suite
12 typed AI ops: ModeSelectOp, AIBoolOp, AIScoreOp,
AIParseNumberOp, AIExtractStringSliceOp, AIExtractMapOp,
AISummarizeOp, AIComputeStringToStringOp, AIBestMatchOp,
AIRerankOp, AIClassifyMultiLabelOp, and
AIComputeMathOperandsToFloat64Op. All target claude-sonnet-4-6 with
up to 3× retry on parse failure.
MCP integration
MCPCallOp and MCPScriptOp[In, Out] for single calls and
multi-call sessions sharing server state. Warm-replenish pool (pool_size,
pool_prewarm, library.ShutdownMCPPool), with
MCPArgsFormatter and MCPResponseParser interfaces for
custom JSON shapes and MCPToolError for tool-level errors. See
the MCP guide.
Retrieval & citations
library.Retriever interface with RetrieveOp /
RetrieveWithFiltersOp. Citation validation via deterministic
ValidateCitationsOp against the retrieved allow-list.
EmbeddingClient / EmbeddingClientFactory abstractions for
vector-store backends; ResolveEmbeddingClient threads credentials
through ctx. See the retrieval guide.
AI-driven repair
library.WithRepair[Inner] wraps any deterministic op with LLM-assisted
recovery when it returns *library.ErrRepairable. Bounded by
max_attempts; input types opt in by implementing
UnmarshalRepair. Repair traces captured via
library.WithReasoningLog. See the repair guide.
Pluggable credential routing
AIClientFactory interface for sourcing Anthropic / Gemini clients from
Vault, AWS Secrets Manager, workload identity, per-tenant credential vaults.
SetDefaultAIClientFactory, RegisterAIClientFactory(id, ...),
and per-vertex client_factory_id + credential_ref params.
See the credential-routing guide.
Examples
Self-contained example programs spanning fan-out/coalesce, parallel AI extractors, MapOver, and multi-label classification, plus Local MCP — Playwright fan-out, Remote MCP — Cloudflare docs, Stock Analyzer (zero custom ops), RAG (BM25), RAG (Gemini embeddings), and WithRepair (JSON + XML).