PR review bot
GitHub webhook → agent reads the diff, tests the coverage gap, emits structured findings back as inline review comments.
--template pr-review
v0.1.0 · Shipped on npm
A stable runtime core. Declarative configs next to the code they automate. Fleets of agents that deploy together. One manifest, one dev loop, one history.
$ npm i -g
@declaragent/cli
Star on GitHub
Use cases
Every row ships as a starter template. declaragent init --template <name>
scaffolds the agent, wires its source, and drops the skill definitions — you fill in
the system prompt.
GitHub webhook → agent reads the diff, tests the coverage gap, emits structured findings back as inline review comments.
--template pr-review
Socket-mode Slack app → agent greps + reads your repo → answers in-thread with file paths and snippets. No RAG index to maintain.
--template concierge
Alertmanager fires → agent grades severity, pulls the runbook, pages the right oncall engineer over Slack DM with a pre-drafted mitigation.
--template oncall-escalator
Kafka in, enriched event out. DLQ, retry backoff, idempotency cache, and a daily token budget are all declarative — no handwritten consumer loop.
--template kafka-pipeline
A concierge that routes to a pr-reviewer, a security-scanner, and a doc-writer over agent-rpc. One manifest, one dev loop, one atomic deploy.
--template fleet-starter
One daemon, many tenants. Per-tenant quotas, extensions, secret scopes, and a hash-chained audit log you can pipe into any SIEM.
--template multi-tenant-starter
Try it — no install
fleet.yaml. Run the actual validator in your browser.
This is the same validation logic from @declaragent/cli, ported
verbatim to the page. Break the YAML — dangle a peer, duplicate a capability,
reference a missing deploy target — and the findings show up below.
How it works
One manifest. One source of truth. No hidden state in a web UI.
version: 1
name: acme-fleet
agents:
- id: concierge
path: ./agents/concierge
- id: pr-reviewer
path: ./agents/pr-reviewer
deploy:
target: cloud-run
deploy:
targets:
cloud-run:
kind: gcp-cloud-run
region: us-central1
One process, shared in-memory bus, inter-agent RPC for free.
$ declaragent fleet run
fleet: acme-fleet
running 2 agents:
• concierge env=default (client-only)
• pr-reviewer env=default capabilities=1
ready. press ctrl-c to stop.
Rolling, all-or-nothing, or per-agent. Rollback is one flag away.
$ declaragent fleet deploy --target cloud-run
[01/02] deploying concierge@v1.2.0-a1b2c3d ✓
[02/02] deploying pr-reviewer@v1.2.0-a1b2c3d ✓
recorded to .declaragent/fleet-deploys.jsonl
Why this
git blame agent.yaml tells you when the agent got slower,
quieter, or stricter. git revert rolls it back. Your agents
have a changelog.
One fleet.yaml declares N agents + shared peers + one deploy.
Rolling or all-or-nothing, both with health-gated rollback. No ad-hoc
orchestration glue.
A typed envelope, pluggable transport, RequestAgent in the
tool box. Kafka / NATS / SQS / AMQP / MQTT in prod; in-memory in dev. Same
wire format.
Apache-2.0, on npm. No SaaS tier gates a feature. The CLI you install is the same binary we run.
curl -fsSL https://declaragent.dev/install.sh | sh
npm i -g @declaragent/cli
brew tap declaragent/tap
brew install declaragent
Then:
declaragent init --fleet my-fleet
cd my-fleet
declaragent fleet add --template rpc-server
declaragent fleet run