Skip to content

graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs

Python
3.0 KB
Raw Blame

What the harness comparisons mean for an engineering budget

A rough guide built from two measured comparisons, each on the same model and account in both harnesses:

Both ran the same 21 coding and MCP tasks, 3 times each. Costs are list prices of the tokens each harness used.

Per task

Compared with graff Other harness graff costs
Codex app server, its reported usage $0.0131 $0.0426 69% less
Codex app server, with its unreported warm-up request $0.0131 up to $0.0672 up to 81% less
Grok CLI, grok-4.7 public list price $0.039 $0.26 to $0.33 85% to 88% less
Grok CLI, xAI's price for its build-model route $0.039 $0.098 60% less

At company scale

An illustrative organization: 100 engineers, each running 40 agent tasks per working day, 21 working days a month, so 84,000 tasks a month at the per-task costs above.

Compared with graff per month Other per month Saved per month Saved per year
Codex app server, reported usage $1,100 $3,578 $2,478 $29,736
Codex app server, with warm-up $1,100 $5,645 $4,544 $54,533
Grok CLI, public list price $3,284 $22,042 to $27,770 $18,757 to $24,486 $225,086 to $293,832
Grok CLI, xAI's build-model price $3,284 $8,240 $4,956 $59,472

Time moves too. Against the Grok CLI, graff took 15.5s per task against 106.9s: about 2,100 fewer hours of agent wall time a month at this volume. Against the Codex app server graff is now faster, 22.2s against 28.6s per task (about 149 fewer hours a month).

How far this carries

  • Real tasks are bigger. These tasks average about 30k prompt tokens in graff. Cost grows with tokens, so the percentage is the part that transfers: an organization spending $X a month on agent runs in one of these harnesses would spend roughly the "graff costs" share of $X for the same work.
  • MCP-heavy work saves the most. Most of the gap against the Grok CLI is on MCP tasks, where it re-sends every tool result on every turn.
  • Plans are not per-token bills. Both comparisons ran on subscription sign-ins; the dollars are list-price equivalents, which is what API-key usage pays.
  • Small samples. 3 runs per task on one machine and one account per comparison. Run graff-evals on your own tasks before budgeting from these numbers.

Enterprise

graff is licensed under a modified AGPL-3.0 (LICENSE). From the release after v0.0.302.22, an organization (counting its affiliates) that has raised more than US$500k, has a net worth or valuation above US$500k, or has gross revenue above US$100k a year needs a commercial license for any use, including on its own machines, CI and network. Contact [email protected] for a license, support, and deployment help.