Your procedures,run by an agent,signed off by you
Upload the SOP your team already follows. SOPilot turns it into an agent that handles the routine work in your real systems, while every approval your policy requires still goes to a person. Each action is recorded against the line of the SOP that called for it.
Starts from documents you already have. No prompts to write.
“Compliance must approve any order over $500, any supplier with adverse findings, or any supplier rated below 3.5.”
- Policy check: Rating 2.9 is below 3.5 → needs approval
- Compliance rejected the vendor. Run stopped at step 7.
You own the agent
Every compile produces a bundle of plain files (SKILL.md, mcp.json, policy.yaml) that runs in any Agent Skills and MCP client.
Not agent platforms, where the agent lives in their account and leaves when you do.
Stays in your cloud
Run it on the model provider you already approved, such as Bedrock, Vertex or Azure. No new vendor in your data path, no new security review.
Not hosted agents that send your data through one more company.
Each step gets only its tools
The agent sees just the one or two tools a step needs, and that scoping travels with the bundle.
Not agents handed every tool at once, which is how they end up picking the wrong one.
Turn any written SOP into a working agent
Upload the procedure your team already follows. SOPilot reads it and turns each step into an action, a decision or an approval, with no prompts to write and no workflow builder to learn. The document you already approved becomes the thing that runs.
Make every threshold in your policy a hard stop
Dollar limits, rating cutoffs and vendor checks in your SOP become rules the agent can't override. With a $500 limit, a $501 order waits for Compliance every time, and if a number is missing, it waits too. Your controls hold even when nobody is watching.
Find out how much of any SOP can be automated
Upload a procedure and get a step-by-step breakdown: what the agent can run today, what still needs a person, and why. It's a starting point for deciding which processes to hand off first, and what tools would close the rest.
Route approvals to the right person, with the evidence attached
When a purchase needs Compliance or Finance, the request goes straight to their queue with the supplier rating, the news coverage and the rule that flagged it. They can decide in one click instead of chasing down the details, and nothing is paid until they do.
Trace every action back to the policy that required it
Every step the agent takes is recorded against the line of the SOP it came from, along with what it checked, what it found, what it cost and how long it took. When an auditor asks why a vendor was paid, you can show them the record.
Edit the document to change how the agent works
Raise the approval limit from $500 to $750 in the SOP and the agent follows the new rule on its next run. You can see exactly what changed and when, and policy owners can make the change themselves without a request to engineering.
Take the agent with you
Every compile also writes the agent out as three plain files. Read them, diff them, check them into git, or run them in any Agent Skills and MCP client.
proc-014-new-vendor-onboarding/
- SKILL.md
---name: new-vendor-onboardingdescription: PROC-014. Vet a new vendor and place the first order.--- ## Step 1 · Collect quotes (§5.2)Collect 3 price quotes from sellersrated 4.0 or higher. Pick the lowesttotal, shipping and tax included. > "Obtain at least three quotes from> sellers rated 4.0 or higher..." ## Step 7 · Compliance approval (§5.6)Stop. Do not pay until Complianceapproves.SKILL.md · Each step in plain English, with the SOP line it came from.
- mcp.json
{ "mcpServers": { "serpapi": { "command": "uvx", "args": ["sopilot-tools", "serpapi"], "env": { "SERPAPI_KEY": "${SERPAPI_KEY}" } }, "nessie": { "command": "uvx", "args": ["sopilot-tools", "nessie"], "env": { "NESSIE_KEY": "${NESSIE_KEY}" } } }}mcp.json · The tool servers it connects to: live web search and banking.
- policy.yaml
steps: 1: { tools: [serpapi.google_shopping] } 2: { tools: [serpapi.google_maps] } 3: { tools: [serpapi.google_news] } 9: { tools: [nessie.transfer] } 10: { tools: [nessie.purchase] } approvals: - step: 7 approver: compliance when: > order_total > 500 or len(adverse_findings) > 0 or rating < 3.5 # A value the rule needs is missing?# Ask the approver. Never guess.on_missing: require_approvalpolicy.yaml · Which tools each step may touch, and the rules that stop it.
Your keys stay in your environment. The bundle only names them.
Built on tools you already trust
Compiler
Gemini
Reads your written SOP and turns it into steps, rules and approvals the agent can run.
Web lookups
SerpApi
Pulls live prices, ratings and news so every vendor gets screened before money moves.
Payments
Capital One Nessie
Checks account balances and pays vendors once the policy and the approver say yes.
Trace store
Tiger Data
Every step lands in a Timescale hypertable, so each run has a timeline you can audit.
Runtime
FastAPI
Runs the agent loop, the policy gate and the approval queue behind one typed API.
Hosting
Vultr
The whole stack ships as Docker Compose behind Caddy on a single Vultr box.
Your team wrote the procedure. Let it do the work.
Drop in an SOPA PDF, Markdown or text file. You'll see every step it found, what the agent can run and what still needs a person, before anything runs.PDF.md.txtNo SOP handy? Start from one of our demos, or compile them all at once.
Compile them all