blyte documentation#
blyte is the implementer you put under a frontier model. Plan with Claude Code or another frontier harness, then hand each well-specified task to blyte running on a cheap model. blyte works in its own git worktree, your check command decides when the task is done, and the frontier model reviews the result and merges it.
flowchart LR
P["<b>Frontier model</b><br/>(Claude Code)<br/>plan: split into tasks<br/>with disjoint files"]
subgraph par ["blyte dispatch, in parallel: one worktree and one gate each"]
direction TB
T1["<b>T1</b> · .wt/T1<br/>blyte (cheap model)<br/>implement, commit"] --> G1{{"gate"}}
T2["<b>T2</b> · .wt/T2<br/>blyte (cheap model)<br/>implement, commit"] --> G2{{"gate"}}
T3["<b>T3</b> · .wt/T3<br/>blyte (cheap model)<br/>implement, commit"] --> G3{{"gate"}}
G1 -. "fail: retry" .-> T1
G2 -. "fail: retry" .-> T2
G3 -. "fail: retry" .-> T3
end
P -- "task + tests" --> T1
P -- "task + tests" --> T2
P -- "task + tests" --> T3
G1 -- "JSON summary" --> R["<b>Frontier model</b><br/>review each diff<br/>as it lands"]
G2 -- "JSON summary" --> R
G3 -- "JSON summary" --> R
R -- "merge one by one,<br/>rebase if main moved" --> M[("main")]
Set up#
- Install blyte:
cargo install blyte --locked - Give it a cheap model:
export GEMINI_API_KEY=...(or another provider). - Install the Claude Code skill:
blyte skill install. blyte becomes Claude Code's default implementer. - Ask Claude Code for a change as usual. It plans, dispatches independent tasks to blyte in parallel, reviews each result and merges.
Getting started walks through each step, and how to run blyte on its own.
Find what you need#
blyte dashboard: a local page of every running and recent job, with time, cost and tags.
Choose the implementer modelWhat has been measured, the six providers, credentials, retries, cost.
Know what commands can touchThe sandbox on macOS and Linux, credential reads, the environment filter, --network=off.
Read a run's resultsEvery field of --report: outcome, turns, cost, timing, security settings.
Look up a flagThe full blyte --help and the environment variables blyte reads.
Understand the designThe loop, the sealed gate, folded context, the tools and their limits.
Key ideas#
The gate. --gate is a command, not a tool the model can call. When the model says it is
finished, blyte runs every step of the gate; exit 0 is verified, anything else goes back to the
model with all the failures at once. See Running tasks.
Rollback. By default a failed gate resets the tree to the commit the run started from, discarding uncommitted changes. Start from a clean tree or a dedicated worktree.
Budgets. Every run is bounded by turns (--max-turns) and gate attempts (--max-gates), and
optionally by wall-clock time (--max-seconds) and spend (--max-cost).
The sandbox. On by default: commands cannot write outside the workspace, and on macOS cannot read credential stores. See Security.
One job. blyte implements one task per run. Planning, splitting work and review belong to the frontier model or person driving it: see the workflow.
Background#
- Manifesto: the design rules and why they exist.
- Evidence: how blyte has been measured, and the caveats.
- The architecture decision and research notes.