ThumbGateThumbGate Verification evidence
comparison | thumbgate vs dirac

ThumbGate vs Dirac

Dirac is an Apache-2.0 coding agent focused on token cost: hash-anchored edits, AST-native refactoring, multi-file batching, native tools and no MCP. ThumbGate is not a coding agent. It evaluates a tool call before it runs, blocks the unsafe ones, and leaves a reviewable record.

ThumbGate: PreToolUse allow / warn / deny on Claude Code, Cursor, Codex
Dirac: cheaper coding agent (their eval, not ours)

The short answer

Not a clone. Same buyer (people running coding agents). Different job. They answer “make the agent cheaper and more precise.” We answer “before this agent curls a secret or force-pushes, stop it.” Dirac’s public table (fetched 2026-08-19) reports 8/8 tasks at $0.18 average versus $0.53 for other agents and a 64.8% cheaper headline. That is their methodology on gemini-3-flash-preview. ThumbGate does not inherit those numbers.

Token waste is not a substitute for a gate

Sample — not a live buyer run. Mandate: apply a cheap multi-file edit. Do not exfiltrate secrets.

01 Agent proposes an editHash-anchored or AST batching may keep the prompt small.
02 Rule checkedSecret-exfil is a hard floor. Cheap diffs still get evaluated.
03 Human / hook checkpointA PAT in the command is denied before execute.
04 Recorded resultDeny is logged. Token savings do not prove the action was safe.

Comparison table

DimensionDiracThumbGate
What it is Open-source coding agent (Cline-family, Apache 2.0) Local PreToolUse firewall for coding agents
Question it answers Can this refactor land cheaper and still compile? Should this tool call run at all?
Edit mechanic Hash-anchored lines + AST-native batching Does not edit code. Evaluates the proposed tool call.
Default hard blocks Agent policy / user confirmation inside Dirac Secret exfiltration and self-protect; destructive git/fs warn unless STRICT
Approvals Default act asks for approval. dirac -y yolo auto-approves every tool. Hook verdict: allow, warn, or deny. Secret-exfil is a hard floor even when an agent is in yolo.
MCP / intercept Native tools only. They advertise no MCP. MCP + PreToolUse on Claude Code, Cursor, Codex. No --agent dirac. Dirac calls are not intercepted today.
Cost claim 64.8% cheaper on their 8-task public table (2026-08-19 fetch) We do not republish that figure as ours. Context-footprint and reasoning-efficiency guardrails exist for token waste inside our loop.
Commerce Free OSS agent + VS Code / CLI Local evaluate is free. Hosted Enterprise is not generally available.

When to use Dirac

  • You want the coding agent itself to be cheaper (hash-anchored edits, AST batching, lean context).
  • You prefer native tool calling and no MCP in that agent.
  • You are comparing open-source agents on Terminal-Bench-2-style evals.

When to use ThumbGate

  • Claude Code, Cursor, or Codex is about to run Bash, Write, or git on a sensitive repo.
  • You need secret-exfil deny locally, with a log you can show a security owner.
  • An agent’s yolo / auto-approve mode is not a substitute for a hook. Token efficiency is not an allow verdict.

FAQ

Is Dirac a ThumbGate competitor?

Adjacent. They are a coding agent. We are a coding-agent tool-call gate. Do not pitch us as their editor.

Do we claim 64.8% cheaper?

No. That is Dirac’s public eval. Cite dirac.run if you quote it. We do not copy their AST engine.

Can ThumbGate block Dirac tool calls today?

No. npx thumbgate init --agent dirac is not a supported path. Dirac uses native tools and no MCP, so our PreToolUse hooks do not see those calls. Use ThumbGate on Claude Code, Cursor, or Codex. A Dirac adapter is not shipped.