FixClaw

AI automation for business operations -- with governance built in.

FixClaw is a pipeline engine written in Go that puts AI to work on real operations workflows (email triage, customer follow-ups, job classification) while enforcing token budgets, audit trails, input sanitization, and human-in-the-loop approval on every outbound action.

AI never executes. Deterministic code does.

Quickstart

git clone https://github.com/renezander030/fixclaw.git && cd fixclaw
cp secrets.yaml.example secrets.yaml   # add your Slack + API keys
go build -o fixclaw . && ./fixclaw

Define your pipelines in config.yaml, your prompts in skills/, and FixClaw handles the rest.

Why FixClaw

Operations teams use AI assistants for personal productivity. But when AI output touches customers, contracts, or compliance, you need more than a chatbot.

	Claude Dispatch	OpenClaw	FixClaw
Purpose	Personal productivity	Personal AI agent	Business operations
Governance	Anthropic-managed	None	You own it: YAML pipelines, token budgets, audit trail
Human-in-the-loop	Pause on destructive actions	Optional	Every outbound action requires operator approval
Token budgets	None (subscription)	None	Per-step, per-pipeline, per-day limits
Prompt injection defense	Platform-level	None	Input sanitization + output schema validation
Data residency	Anthropic cloud	Self-hosted	Self-hosted. Your data stays on your infrastructure
Configuration	Natural language	Natural language	YAML. Deterministic, version-controlled, auditable

Governance and Compliance

Every pipeline run produces a verifiable audit trail:

• Token budgets -- per-step, per-pipeline, and per-day limits. Exceeding any budget halts the pipeline immediately. No silent overruns.
• Human-in-the-loop -- approval steps present AI output to the operator via Slack/Telegram with approve/edit/reject controls. Nothing leaves the system without explicit sign-off.
• Input sanitization -- operator input is scanned for prompt injection patterns, stripped of role markers and formatting that could break prompt boundaries. Rejected inputs are logged silently (no information leakage to attacker).
• Output validation -- AI output is validated against the skill's JSON schema. Type checks, range enforcement, required fields. Invalid output is rejected.
• Rate limiting -- per-user, per-minute limits on operator interactions prevent abuse.
• Channel security -- allowed user lists, input length limits, and markdown stripping are enforced at startup. The engine refuses to start without security configuration.

How It Works

FixClaw runs pipelines. Each pipeline is a sequence of typed steps:

Step type	What it does
deterministic	Plain code: fetch emails, filter, route, notify
ai	LLM inference with a skill template, budget-checked
approval	Human-in-the-loop: operator reviews before proceeding

Example pipeline:

pipelines:
  - name: email-digest
    schedule: 30m
    steps:
      - name: fetch-unread
        type: deterministic
        action: email_unread

      - name: summarize
        type: ai
        skill: email-digest

      - name: report
        type: deterministic
        action: notify

Architecture

Configuration

config.yaml

Defines LLM providers, models, token budgets, and pipelines.

provider:
  type: openrouter
  api_key_env: OPENROUTER_API_KEY
  base_url: https://openrouter.ai/api/v1

models:
  haiku:
    model: anthropic/claude-haiku-4-5
    max_tokens: 1024
  gpt-4o-mini:
    model: openai/gpt-4o-mini
    max_tokens: 1024

budgets:
  per_step_tokens: 2048
  per_pipeline_tokens: 10000
  per_day_tokens: 100000

secrets.yaml

Private values that stay out of version control. Copy secrets.yaml.example to get started.

Skills

YAML prompt templates in skills/. Each skill defines the system prompt, input variables, and optional output schema for validation.

# skills/classify-job.yaml
name: classify-job
system: |
  You are a job classifier. Given a job posting, determine if it matches
  the freelancer's profile. Return a JSON object with:
  - match: boolean
  - reason: string (one sentence)
  - score: number (0-100)
input_vars:
  - posting
  - profile
output_schema:
  type: object
  required: [match, reason, score]

Project Structure

fixclaw/
  main.go          # Engine: pipeline runner, operator bot, scheduler, guardrails
  gmail.go         # Gmail / Microsoft 365 integration (OAuth 2.0, read + send with HITL approval)
  config.yaml      # Pipelines, models, budgets, timeouts
  secrets.yaml     # Private config (operator IDs) -- gitignored
  skills/          # Prompt templates with schema validation

License

MIT. See LICENSE.