Can axonpush block a tool call?

Yes. Rules run inline in the gateway and can block, redact or flag on the request, the response, or a specific tool call. A blocked tool call never executes; the decision is recorded with the rule that fired.

Every gateway decision: the model, the outcome, and any enforcement action taken (allow, redact, block or flag). The trail is exportable to CSV or JSON so you can hand it to an auditor or feed it into your own system.

Yes. A spend policy sets a USD limit over a window and steps through a ladder of actions as spend climbs: notify your team, fall back to a cheaper model, then block. Scope each policy to an app, environment, model, key, user or tag.

No. Point your existing OpenAI or Anthropic client at the axonpush gateway and add your key. The gateway proxies the call and records it. It is fail-open: if axonpush is unavailable the call still reaches the provider. SDK and OpenTelemetry paths exist if you prefer them.

Will it show every part of my backend?

The gateway sees every call that goes through it. To follow a request across your own services, propagate trace IDs with an SDK or OpenTelemetry. Missing instrumentation means missing spans, not missing gateway calls.

Where does my data go?

Hosted traffic is processed by axonpush. BYOC runs in your own AWS account. Redaction runs inline before a call leaves your network, and optional semantic analysis can send configured content to model providers, so review consent, redaction and provider settings before enabling it.

Observe and control every AI agent. One line of code.

axonpush sits inline on every LLM and agent call. See tool calls and handoffs, block or redact unauthorized actions, cap spend, and keep an audit trail. No SDK, no code changes beyond the base URL.

One base-URL change puts axonpush inline on every call.

Start freeSee the dashboard

Free 14-day Pro trial. No credit card. Change one line to start.

axonpush dashboard with populated activity metrics, an hourly chart and intensity cells, using synthetic demo data

Real product. Synthetic demo data.

Select the image to explore the full dashboard.

See
Every model call, tool call and agent handoff, as one queryable trace.
Control
Block, redact or flag a request, response or tool call. Cap spend, tokens and reasoning. Restrict models and tools.
Prove
An immutable, exportable record of every decision and the action taken.

Works with any OpenAI or Anthropic client.

axonpush speaks the OpenAI and Anthropic wire formats and reads OpenTelemetry, so it drops in front of the SDKs and frameworks you already use.

Frameworks and providers

LangChain
LangGraph
OpenAI
Anthropic
CrewAI
Deep Agents
Vercel AI
Mastra
Google ADK
LlamaIndex
OpenTelemetry
Pino
Winston
loguru
structlog

Agents don't just answer anymore. They act.

An agent calls a tool, issues a refund, reaches for an account, reads a record. The model output looks fine. The action is the risk, and by the time you read the transcript it has already happened.

axonpush watches the action, not just the answer. It sees the tool call before it runs and can stop the ones that break policy.

One inline position, two jobs.

Because axonpush proxies the call, the same place that records what happened can also decide what happens next. Observability and enforcement are one system, not two tools bolted together.

Change one line. Keep shipping.

Point your OpenAI or Anthropic client at the axonpush gateway and add your key. It proxies the call to your provider, records the full request and response, model, tokens, cost and latency, and streams the answer straight back.

Fail-open by design. If axonpush is unavailable the call still reaches the provider. No new SDK, no code changes beyond the base URL.

client = OpenAI(
-  base_url="https://api.openai.com/v1",
+  base_url="https://gateway.axonpush.xyz/v1",
   api_key=OPENAI_API_KEY,  # unchanged
+  default_headers={"x-axonpush-key": AXONPUSH_KEY},
)

Fail-open: if axonpush is unavailable, the call still reaches the provider.

See the failed step, not just the final answer.

Open a trace to see the model calls, tool calls and agent handoffs behind a response, with inputs, outputs, errors and timing together, and the step axonpush enforced marked in place.

A support agent called issue_refundwith amount: 25000. A rule on the tool argument blocked the call before it ran, and the block sits on the trace where it happened.

A closer look.

A blocked $25,000 refund in a synthetic support request.

Detail of a support-agent trace with a model span open on Input / Output: the instruction to issue the refund and the issue_refund tool call it produced, using synthetic demo data

A support agent's blocked $25,000 refund, synthetic demo data.

Slice usage by your own dimensions.

axonpush keeps every attribute you put on a span, so the dimensions you analyse by are not a fixed list. Stamp aparticipant_role,tenant orplan and break down, trend and get latency for any of them, next to the built-in model, provider and tool views.

Because a dashboard is a spec of widgets over that same query surface, your coding agent can read your backend and author one tailored to your product over MCP, no dashboard-building by hand.

Your dimensions, not just ours.

Pick a business attribute your app emits and get its breakdown, trend and latency percentiles in one view, then save it as a dashboard.

axonpush Usage explorer grouped by a custom participant_role dimension, showing the candidate-versus-recruiter breakdown, latency percentiles and a calls-over-time trend, using synthetic demo data

Sliced by a custom business dimension, synthetic demo data.

Block what breaks policy. Cap the spend.

Rules run inline in the gateway, before the call leaves your network. Detect PII, secrets, org-defined patterns, tool names and tool arguments, then redact, block or flag each match. A blocked call never reaches the provider; a redacted one is forwarded with the value removed.

A spend policy sets a USD limit over a window and steps through a ladder as spend climbs: notify your team, fall back to a cheaper model, then block. Scope each policy to an app, environment, model, key, user or tag.

Enforcement you can see.

The violations feed shows which rules fire and how often. The spend bar shows each policy against its cap, and exactly what fires at 80% and 100%.

axonpush moderation violations feed showing a blocked large-transfer tool call and the top violated rules, using synthetic demo data

Violations caught inline, with the rule and action for each, synthetic demo data.

axonpush spend policies page showing USD limits, spend bars near their caps and notify-then-block ladders, using synthetic demo data

Spend policies climbing a notify-then-block ladder, synthetic demo data.

Bound the request itself.

Moderation watches content and spend policies watch cost. Govern policies bound the request a caller may send: cap output and reasoning budgets, restrict which models can run, pin the service tier, allow or deny tools, and require structured output or an end-user id.

Choose how each policy reacts. Enforce rewrites the request to comply and forwards it, block rejects it, warn records the would-be change. Scope each to an app, environment, model, key or user.

Guardrails on every call.

A capped max_tokens, a stripped tool, a model rewritten to the approved one, each recorded on the call so you can prove exactly what axonpush changed.

axonpush governance policies page showing request-mutation rules that cap token and reasoning budgets, restrict models, pin the service tier and allow or deny tools, using synthetic demo data

Govern policies bounding the request, enforced, blocked or warned, synthetic demo data.

Prove what happened.

Every gateway decision is recorded: the model, the outcome, and the action taken. Allow, redact, block or flag, with the rule that fired. The decision trail is exportable to CSV or JSON for an auditor or your own system.

An immutable record.

Continuous runtime evidence of what your agents did and why, not a point-in-time snapshot.

axonpush decision trail showing per-decision allow, redact, block and flag outcomes with the model and rule for each, using synthetic demo data

The decision trail, exportable to CSV or JSON, synthetic demo data.

Get ahead of the EU AI Act.

The EU AI Act's high-risk obligations are phasing in through 2026 and 2027, and they reach well beyond finance. axonpush gives you the runtime evidence those rules ask for.

The decision trail is the record-keeping and logging you can point to for Article 12. Block, redact, flag and escalate-to-human are the human oversight controls behind Article 14.

Evidence that helps you meet these obligations. Not legal advice or a guarantee of compliance.

Maps to

EU AI Actlogging (Art. 12) · human oversight (Art. 14)
DORAICT risk monitoring and incident evidence
SR 11-7model-use records for risk management
SOC 2controls in progress; ask about current status

Wherever your agents act.

The same trace, control and audit surface, seen from seven places an agent can go wrong. Synthetic demo data throughout.

Select a card for the full story.

Start with the stack you already use.

Point your base URL at the gateway, connect an existing OpenTelemetry exporter, or add the SDK to your app. Prefer working with a coding assistant? Give it the setup prompt.

client = OpenAI(
-  base_url="https://api.openai.com/v1",
+  base_url="https://gateway.axonpush.xyz/v1",
   api_key=OPENAI_API_KEY,  # unchanged
+  default_headers={"x-axonpush-key": AXONPUSH_KEY},
)

Start hosted. If your team needs its own AWS deployment, we will walk through BYOC licensing, setup and data-processing requirements. AWS usage charges apply.

Explore AWS deployment

Ways in

Point your base URL at the gatewayno SDK, no code changes
OpenTelemetryconnect an existing OTLP exporter
pip install axonpushPython SDK
npm i @axonpush/sdkTypeScript SDK

Hosted or your AWS account

Hostedstart free, upgrade when you need to
Your AWS accountlicensed BYOC deployment
Serverless analyticsno analytics cluster to manage
Explicit data controlsreview redaction and model-provider settings

Know what you are connecting.

What you can block, what gets audited, how spend is capped, and where your data goes.

Can axonpush block a tool call?: Yes. Rules run inline in the gateway and can block, redact or flag on the request, the response, or a specific tool call. A blocked tool call never executes; the decision is recorded with the rule that fired.
What gets audited?: Every gateway decision: the model, the outcome, and any enforcement action taken (allow, redact, block or flag). The trail is exportable to CSV or JSON so you can hand it to an auditor or feed it into your own system.
Can I cap spend?: Yes. A spend policy sets a USD limit over a window and steps through a ladder of actions as spend climbs: notify your team, fall back to a cheaper model, then block. Scope each policy to an app, environment, model, key, user or tag.
Does it need an SDK?: No. Point your existing OpenAI or Anthropic client at the axonpush gateway and add your key. The gateway proxies the call and records it. It is fail-open: if axonpush is unavailable the call still reaches the provider. SDK and OpenTelemetry paths exist if you prefer them.
Will it show every part of my backend?: The gateway sees every call that goes through it. To follow a request across your own services, propagate trace IDs with an SDK or OpenTelemetry. Missing instrumentation means missing spans, not missing gateway calls.
Where does my data go?: Hosted traffic is processed by axonpush. BYOC runs in your own AWS account. Redaction runs inline before a call leaves your network, and optional semantic analysis can send configured content to model providers, so review consent, redaction and provider settings before enabling it.

See, control, and prove what your agents do.

Point one client at the gateway. Watch the calls, block what breaks policy, and keep the record.

Start free See pricing

No card required to start.

Want a hand wiring it in?

Bring an agent and one tool call worth governing. Work through the setup directly with the founder.

Book a setup call

Observe and control every AI agent. One line of code.

See

Control

Prove