toolbench

toolbench is a platform and CLI for building benchmarks for agentic tools and harnesses. Compose a benchmark (a task + a grading rubric) with a harness (the agent runtime), a loadout (the tools the agent is given), and a variant (the prompt + sandbox), run N trials per cell against a model, and get reach / pass@k / pass^k metrics with plots.

It is the benchmarking sibling of toolbase — the package manager and runtime for agent tools. toolbase serves the tools; toolbench measures how well an agent uses them. Together they close the loop for agentic tool and harness development.

Install

pip install toolbench                 # core framework
pip install 'toolbench[toolbase]'     # + resolve tools from toolbase loadouts

Requires Python ≥ 3.12. The CLI is available as both toolbench and the short alias tbe.

Quickstart

The example geometry benchmark (Euclidean distance + midpoint between two 2-D points) is a self-contained, dependency-free example that exercises the whole framework.

# Validate the wiring with no LLM calls or cost:
toolbench run --benchmark examples/geometry --model stub \
    --loadouts full_local --n 1 --max-cost-usd 0 --dry-run

# A real run (a few cheap trials):
toolbench run --benchmark examples/geometry --model claude-haiku-4-5 \
    --loadouts core_only,full_local --n 3 --max-cost-usd 0.50

Each run writes a directory under runs/<run_id>/ with a manifest.json, per-trial transcripts and artifacts, an aggregated summary.json / summary.txt, and headline plots (k-sweep, parallel-coordinates, per-stage breakdown).

Concepts

A benchmark lives in examples/<name>/ and is composed from four declarative axes — vary any of them on the command line to run an ablation:

A loadout source is one of:

• python: — import a module exposing TOOLS / make_tools() (the no-dependency escape hatch),

• toolbase: — resolve tools from a toolbase loadout, in-process (requires toolbench[toolbase]); served toolkit versions are recorded in the run manifest as reproducibility provenance, or

• mcp: — serve any MCP server's tools, stdio or HTTP (requires toolbench[mcp]):

  tools:
    sources:
      - toolbase: { loadout: my-loadout }
        select: [calculator__add]          # optional: ablate within the loadout
      - mcp: { url: "https://host/mcp", headers: { Authorization: "Bearer ${TOK}" } }
  

Metrics

For each (model × condition) cell over k trials:

• $\overline{\text{reach}}_k$ — mean rubric-weighted reach: how far through the task the agent gets, on average.
• $\text{pass@}k$ — probability that at least one of $k$ trials passes (best-of-$k$).
• $\text{pass}^{k}$ — probability that all $k$ trials pass (worst-of-$k$).

with bootstrap 95% confidence intervals and a metric-correlation matrix. A trial passes when it clears the rubric's pass criterion — every stage by default, or reach ≥ a pass_threshold once the rubric uses partial-credit stages. See Metrics for the exact estimators.

Commands

Run toolbench --help (or tbe --help) for the full reference.

Also

• Runtimes — besides the API-driven orchestral runtime, toolbench drives the claude_code and codex CLIs directly (subscription auth, no API key), with per-turn token accounting and a filesystem sandbox. Copy-paste starting points live in harness_templates/.
• Judges — grade with the deterministic rule judge (default) or add an LLM second opinion (--judge rule+llm); the rule grade always stays authoritative. Any judge can be applied after the fact with toolbench regrade --judge ….
• Safeguards — trials that read the ground-truth answer key are quarantined (INTEGRITY_LEAK, scored 0) so a leak can't inflate the headline; every trial also gets a readable audit.txt of its full trajectory.

License

MIT