-
Notifications
You must be signed in to change notification settings - Fork 0
Comparison
PromptPilot is closest to an SLM-powered harness, not a pure token reducer or coding agent.
| Category | Main idea | PromptPilot difference |
|---|---|---|
| Token reducer | Make context smaller | PromptPilot uses an SLM to decide when compression is safe |
| Prompt optimizer | Rewrite prompts | PromptPilot can clarify, answer, pass through, or invoke the agent |
| Model router | Choose cheap or expensive model | PromptPilot routes workflow actions, not only model calls |
| Agent orchestrator | Run agents | PromptPilot controls the context and decisions around agent execution |
| Coding agent | Write/debug code | PromptPilot delegates coding to Codex/Claude-style agents |
| Third-party Claude harnesses (OpenClaw, OpenCode) | Extract Claude OAuth tokens and make direct API calls | PromptPilot does not handle the token at all — see Compliance posture below |
PromptPilot's core claim is bounded: a small model can manage the workflow around a large coding model. The small model should not be judged as a replacement coder; it should be judged as a harness for routing, clarification, semantic preservation, and safe context control.
The April 2026 enforcement action against OpenClaw and OpenCode targeted
tools that read OAuth credentials out of ~/.claude/.credentials.json and made
direct calls to api.anthropic.com while spoofing Claude Code's request shape.
That violates Anthropic's Feb 20, 2026 clarification: subscription OAuth
credentials are scoped to "ordinary use of Claude Code and other native
Anthropic applications," and the literal prohibition is on using those tokens
in any other product, tool, or service.
PromptPilot is structurally different. When you run prpt against the Max
subscription, PromptPilot invokes the official claude binary as a child
process and sends it a rewritten prompt. The credential never enters our
process; from Anthropic's server logs the request originates from the real
claude binary with its real user-agent and request shape, because that's
what made the call.
| Axis | OpenClaw / OpenCode | PromptPilot |
|---|---|---|
| OAuth token handling | Read ~/.claude/.credentials.json; sent the bearer directly to the API |
Never reads, stores, or transmits the token; lives inside the official claude binary process |
| API endpoint | Direct HTTPS to api.anthropic.com/v1/messages
|
No direct Anthropic API calls from our process; the official binary makes them |
| Request shape | Spoofed Claude Code's headers, system prompt, and tool definitions | The real claude binary builds the request with its real shape |
| Client identity | Client replacement — pretended to be Claude Code | A local user invoking their own already-logged-in CLI from a parent process |
| Fan-out | Shared service letting many users plug accounts in | Local process, one user, one machine |
| Orchestration vs replacement | Replaced Claude Code with a third-party client | Drives the official Claude Code with a rewritten prompt |
The literal token-clause prohibition does not apply to PromptPilot's
subprocess pattern. What remains interpretive is the broader "ordinary use"
phrasing in the Feb 20 statement — whether driving claude -p programmatically
counts as "ordinary use" depends on Anthropic's interpretation, and the answer
is probably "yes for modest interactive use, ambiguous for high-volume
automation." For sustained automation (multi-hundred-call experiments, CI
loops, etc.) the conservative path is ANTHROPIC_API_KEY with the SDK
normalizer, which is documented as a first-class option in
Authentication and Providers.
PromptPilot prints a once-per-process informational note when the subscription
path is used, so users can make their own call about which authentication mode
fits their workload. The same reasoning applies to ChatGPT/Codex subscription
routing via codex; OpenAI has not published a Feb-2026-style clarification,
but the structural pattern (subprocess of the official binary, no token
handling) is the same.
A separate monthly credit pool for the Agent SDK and claude -p is
expected June 15, 2026 — that launch may formalize programmatic
subscription use under a sanctioned billing model.
- You already use Codex or Claude-style coding agents.
- You want fewer wasted frontier-model calls on ambiguous or simple requests.
- You want noisy tool output compressed without losing debugging facts.
- You care about preserving explicit constraints more than maximizing token reduction.
- You want a standalone coding agent.
- You want the small model to make implementation decisions.
- You need fully automatic context deletion with no passthrough fallback.
See also: Project Overview · SLM Harness · FAQ