秉火前行。 — Carry the fire forward.
One person running coding agents — Claude Code, Codex, Grok — who wants their own rules
enforced across all of them. Local and personal: one binary, plain files under ~/.flint.
No server, no daemon, no account, no telemetry.
Requires Rust 1.85+. There is no published release channel; the git tree is the distribution.
cargo install --git https://github.com/PunkGo/flint flint-cliOr paste this to your coding agent, which does the whole thing:
Read https://github.com/PunkGo/flint/blob/main/SETUP.md and set Flint up on this
machine. Follow it exactly. Stop at the signing step and hand me that command —
signing is mine.
SETUP.md is the full procedure. Rules take effect only when you sign
them; an agent never signs. Setup merges a hook into the config of each agent you name
(~/.claude/settings.json, ~/.codex/hooks.json, ~/.grok/hooks/flint.json), backing up
each first.
Signed markdown in ~/.flint/canon/rules/ → flint compile renders per-agent hook
wiring → the agent's PreToolUse hook runs flint hook, which verifies the signature,
matches the call, and returns warn (proceeds), critique (blocked, recovery path) or
block (hard freeze) → a receipt lands in a local log: rule id, verdict, timestamp, never
the command itself.
| Rule kind | Claude Code | Codex | Grok |
|---|---|---|---|
command — regex over a shell command |
blocks | blocks | blocks |
path — glob over a write target |
blocks | advisory only | blocks |
advisory — guidance into agent context |
compiled in | compiled in | not compiled |
A path rule on Codex is guidance, not a gate; Grok never receives advisory rules. The
command row blocks everywhere. Which envelope each harness honours, and when that was
measured, is in ARCHITECTURE.md.
A receipt records Flint's judgment, not the harness's obedience — the two came apart for two weeks on one machine, with identical-looking receipts. Enforcement is proven only by watching a command fail to run, which is why setup ends with a live-fire probe.
flint law list # rules in the working tree, and their status
flint law accept --name <id> # sign one in (yours to run, not your agent's)
flint canon list # the signed set actually in force
flint canon pick # re-sign after editing a rule
flint pit mark "<gist>" # capture a wall as you hit it
flint knowledge review # triage captures; nothing is ever auto-promoted
flint memory capture "<note>" # optional vault, if wired with --with-memory
flint fleet add --pubkey <k> # trust another machine's public keyAll take --config ~/.flint/flint.toml. Fields and config keys:
docs/reference.md.
Public at v0.1.3. CI covers Linux, macOS and Windows. Claude Code and Grok enforcement is live; the Codex adapter is partial. In daily use on the author's machines for months; as of this release nobody else has installed it, and the install path has been walked by hand only on macOS. Most useful thing to report back: whether your harness actually refused to run the command.
SETUP.md— install and operate, written for the agent doing itARCHITECTURE.md— internals, boundaries, per-harness measurementsdocs/reference.md— every rule field andflint.tomlkeydocs/whitepaper.md— why this existsexamples/laws/— sample rules to copy, read, and sign
Threat boundaries, and what Flint explicitly cannot defend, are in
SECURITY.md.
Build, test and guardrails are in CONTRIBUTING.md.
MIT © PunkGo