Rust-native agentic platform engineering with embedded MCP

Demo code for a demonstration of Rust-native agentic platform engineering with embedded MCP.

StackQL treats cloud and SaaS providers as data sources accessed via SQL. Agents doing platform engineering, SRE and audit work need to query, reason about and act on actual running state, not on state files. The StackQL MCP server is the agent interface to that engine, with a small fixed tool surface and safety modes gating writes.

The stackql-mcp crate embeds that server in a Rust application, either as a sidecar (downloaded and verified at first run) or vendored straight into the Rust application (one self-contained binary). The worked examples are two small agents with the same four parts, model, prompts, context and MCP tools: an SRE assurance sweep over a service's footprint in AWS and Cloudflare (Claude, server as a sidecar) and a FinOps report from Cost Explorer (GPT-5, server vendored into the binary).

Quick start

git clone https://github.com/stackql/rust-embedded-mcp-with-stackql
cd rust-embedded-mcp-with-stackql
cp .env.example .env            # AWS, Cloudflare, Anthropic and OpenAI keys for the main flow
set -a; . ./.env; set +a
for p in aws awscc cloudflare github; do stackql exec "REGISTRY PULL $p"; done
(cd embedded && cargo build --release)    # both agents; the vendored one fetches the server bundle at build time

Needs stackql, stackql-deploy, cargo 1.88+ on PATH; psql, jq, node and python for the act 1 extras.

Then, from the repo root:

set -a; . ./.env; set +a
stackql exec -i primer/queries/exposure.iql                        # act 1 (paste blocks in primer/shell.iql, exec.sh, srv.sh)
stackql-deploy build stacks/service-footprint dev --env-file .env  # act 2: SG + instance + DNS record, about 30 s
./embedded/target/release/sre-agent-sidecar --check                # act 3: embedded server, providers, tools, no model call
./embedded/target/release/sre-agent-sidecar                        # act 3: the SRE sweep (Claude, sidecar server)
./stacks/service-footprint/drift.sh && ./embedded/target/release/sre-agent-sidecar   # act 3: drift on both planes, found and explained
./embedded/target/release/finops-agent-vendored                    # act 3: the FinOps report (GPT-5, vendored server)

Why this shape

• One engine, several consumption patterns: shell, exec, Postgres wire protocol, declarative deploy, MCP. Act 1 and act 2 show the first four; act 3 is the fifth.
• The agent's tools are the StackQL MCP tools, so what the agent decides to do is always a readable SQL statement. The tool-call trace is the audit trail.
• Safety is a server-side contract. Mode::ReadOnly refuses writes however the model is prompted; both agents run in it. Mode::Safe asks a human before each write over MCP elicitation (Mode::DeleteSafe asks only for deletes).
• Agency is a property of the process, not the prompt: the mode is set in code, the prompt names what to read, the audit log records every tool call.
• Sidecar and vendored are the same builder with one line of difference. finops-agent-vendored is a single file with the StackQL engine inside that starts with no network.
• The prompts are markdown compiled into the binary (include_str!), so the persona, the StackQL context and the task are data you can read and edit without touching the Rust.

References

• Crate and docs: stackql-mcp on crates.io, docs.rs/stackql-mcp; source lives in stackql/stackql packaging/mcpb/cargo. The crate version equals the stackql release it embeds - pin the minor (stackql-mcp = "0.10").
• StackQL MCP server: stackql.io/docs/mcp
• stackql-deploy: stackql-deploy.io, stackql/stackql-deploy-rs
• Server bundles: built by stackql/stackql packaging/mcpb, published on stackql/stackql releases
• Provider registry: registry.stackql.app

MIT licensed. Stars, issues and pull requests here and on stackql/stackql are the ask at the end of the talk.