For teams letting AI agents touch production.

Your agent just sent that email twice.

Praxis would have stopped it. Praxis is the one path every action your agents take to the outside world — Slack, SMTP, GitHub, your own APIs. Preview before running. Block what your policy disallows. Never fire the same call twice. Replay exactly what happened later.

Run the demo ★ on GitHub See it work ↓

Stop double-sendsSame email, same Slack, same Jira ticket — Praxis collapses retries before they fire.

Dry-run anythingPreview your agent's next 10 actions before you let it run them.

One audit, one query"Show me every action this agent took last quarter" — answered from one log, not five vendor dashboards.

Scenario

Discover, dry-run, execute, replay, idempotent retry.

No SMTP/Slack credentials required — degraded handlers return a simulated success so the loop completes on a laptop. The load-bearing demo: every step persists to the audit log, and praxis log show reconstructs the entire lifecycle from those rows alone. 

$ make build
$ ./bin/praxis caps list
NAME              IDEMPOTENT  SIMULATABLE
send_email        true        true
send_message      true        true
http_request      true        true
github_create_issue ...

$ ./bin/praxis run send_email \
    '{"to":"alex@example.com","subject":"hi","body":"hello"}' --dry-run
{ "action_id": "act_7a1", "status": "simulated",
  "preview": { "to": "alex@…", "subject": "hi", "body_chars": 5 } }

$ ./bin/praxis run send_email \
    '{"to":"alex@example.com","subject":"hi","body":"hello"}'
{ "action_id": "act_7b2", "status": "succeeded", "external_id": "" }

$ ./bin/praxis log show act_7b2
received    -> validated -> policy.allow -> executed -> succeeded
inputs:     to=alex@example.com subject=hi
policy:     rule=default decision=allow
duration:   12ms

$ ./bin/praxis run send_email \
    '{"to":"alex@example.com","subject":"hi","body":"hello"}' \
    --idempotency-key act_7b2
{ "action_id": "act_7b2", "status": "succeeded", "from_cache": true }
# second call short-circuits — no double effects, recorded in audit
Replay-from-audit is the feature that pays for the runtime. The same executor serves CLI, HTTP, and MCP. The audit log is structurally complete — given only the rows, you reconstruct the full lifecycle. Idempotency is enforced at the destination via keys forwarded from the caller; retries collapse, side effects don't.

How it differs

Where Praxis fits.

Praxis is not a workflow engine, not an agent framework, not a tool router. It's the layer where the act of doing one thing is governed, simulated, and reconstructible.

Tool What it solves What Praxis adds
Computer Use · OpenAI tool-calling Lets the model emit an action. Names the action, blocks what your policy forbids, dedupes retries, and records what happened — so the model can decide and you can still sleep.
n8n · Zapier · raw webhooks Workflow glue and trigger routing. Preview, idempotency, and audit on every single action — not a DAG you have to maintain.

Architecture

Caller → Executor → vendor + audit + outbox.

One executor. One audit chain. One outbox to Mnemos. CLI / HTTP / MCP all run the same path; the audit can't diverge because there's only one path to diverge from.

Run it

Build. Discover. Dry-run. Execute. Replay.

git clone https://github.com/felixgeelhaar/praxis.git
cd praxis && make build

./bin/praxis caps list
./bin/praxis run send_email '{"to":"a@example.com","subject":"hi","body":"hello"}' --dry-run
./bin/praxis run send_email '{"to":"a@example.com","subject":"hi","body":"hello"}'
./bin/praxis log show <action-id>

HTTP surface (same executor):

./bin/praxis serve --port 8080 &
curl -H 'Authorization: Bearer …' \
     -d @action.json \
     localhost:8080/v1/actions
What's shipped today. Capability registry + JSON Schema validation, policy modes, idempotency keeper, audit log + replay-from-audit canary on every commit. Slack, SMTP, GitHub, Linear, Calendar, http_request handlers — credentialed when env vars are set, degraded-mode simulated otherwise. Sigstore-verified Go-plugin loader. OTel traces. MCP federation that aggregates upstream MCP servers as Praxis capabilities. No one-shot ./demo.sh — the CLI flow above is the demo. DryRun parity isn't 100% across every handler; multi-tenant policy + reversibility are scaffolded, not the headline.

Cognitive stack context

Praxis works alone. Plugs into a four-system stack.

Mnemos remembers, Chronos watches time, Nous decides, Praxis executes. Olymp drives the loop. Each is independently usable.

Mnemos · memory

Evidence-linked claims with point-in-time replay and contradiction detection.

Chronos · time

Pattern detection over time-series data. Typed signals, not threshold alerts.

Nous · decisions

Commitment extraction, risk evaluation, intervention generation.

Olymp · runtime

Drives observe → understand → decide → act → learn end-to-end.