Skip to main content
Built-in tools are application code. The lean release has no dynamic runtime integrations.

Definition

Fella’s standard registry contains seven built-in tools. Adding one requires an application code change. Each has a JSON schema and produces an evidence record. The experimental MCP boundary is documented separately, but /mcp is inert in the shipped application. No remote tools can enter the default registry. The extension point is a Rust trait, not a dynamic plugin ABI:
ToolOutput keeps the model-facing text separate from the data shown in the evidence panel:
The standard registry is explicit:
inspect_table is the current combined interface; older commits had separate describe_schema and sample_rows concepts. PR #52 merged that surface so schema statistics and sample rows travel through one tool.

Architecture

The built-ins are offered when a workspace is open. With no workspace, the model receives no data tools. Every successful call is translated into a ToolOutput and then an EvidenceItem with a stable ordered ID for that answer. SQL-backed evidence also maps referenced tables to catalogued files or workbook sheets and retains source-level ingest notes when present. The output has a human-facing summary, compact model text, and optional SQL, columns, rows, text output, or structured chart data. make_chart derives its chart values from its own read-only query, so the chart evidence retains the SQL, result rows, and source mapping; it never accepts or emits SVG/HTML. ChartData is rendered by the Svelte client. Failed calls are also recorded with an error, summary, and elapsed time rather than disappearing from the run.

Flow

For tables, the expected path is inspection when needed followed by run_sql. For documents, the model uses grep_files to locate text and read_file for context; read_file accepts either name or a names array and labels combined documents. make_chart receives a query, executes it through the guarded data path, and derives the chart values itself. The first query column becomes labels; the next one or two become series. run_python provides sql(query), which returns a bounded list of dictionaries from the host’s read-only data engine. SQL and Python are peers: use SQL for relational selection and shaping; use Python when it is the clearer or better-suited method for statistics such as median, quantiles, standard deviation, correlation, regression, and multi-stage numerical work. Combine them when useful rather than contorting one language to cover the other’s strengths. For workspace calculations, use sql(query) inside the Python snippet instead of copying rows from earlier tool output into generated code. The built-in median, stdev, pearsonr, and linregress helpers need no third-party packages; the guest has no pandas, NumPy, SciPy, package installer, or system Python dependency. In particular, pearsonr(x, y) returns one scalar coefficient (not SciPy’s (r, p) result and not a p-value), while linregress(x, y) returns (slope, intercept, r). A correlation coefficient alone does not establish statistical significance.

Behavior

The final answer can include a separate JSON-render composition step. It is not an eighth data tool: it cannot inspect files, run SQL, or create analytical values. It receives accepted result bindings and composes a validated presentation from the registered component catalog. The model selects tools to fit the analytical question: inspect tabular structure and values, query relational data, use Python for statistical work, search/read documents, or combine them when evidence spans formats. No language or tool sequence is the universal route. run_sql accepts a single guarded read-only query, and the SQLite query connection is opened read-only. Direct file-reading SQL functions are blocked. There is no MCP execution path in the personal release. A future connector must preserve the fixed evidence and verification boundary before it can become an official feature.

Limits

  • Python runs in a fresh embedded WASM guest with no filesystem, network, environment, or subprocess capability. Wasmi fuel, memory, stack, source, output, SQL-row, and SQL-response limits bound the calculation; Wasmi, RustPython, and the checked-in guest remain part of the trusted computing base.
  • The Python guest’s fuel, memory, stack, source, output, SQL-row, and SQL-response limits are cross-platform; the host SQL watchdog applies to each SQL call.
  • Tool outputs are capped, so evidence can contain a sample rather than every returned row or character.
Direct /sql is deliberately outside this registry path: it calls the engine directly, prints a system result, and does not create an assistant analysis record or invoke post-answer verification.

Next steps

The harness

Follow tool selection, execution, caching, and stopping.

Experimental MCP

Read the inert extension boundary for future forks.

Verification

See which results can be checked after the answer.

The engine

Understand the local SQL and document paths behind the tools.