Calls

–

Total tool calls compressed

Tokens saved

–

Across all sessions

Avg efficiency

–

Tokens saved per call

CO₂ saved

–

Estimated, from tokens saved

Sessions

–

synthelion-mcp / CLI processes

Avg calls / session

–

Calls / session count

Tools used

–

Distinct tool names

Best single call

–

Highest single-call savings

Avg latency

–

Mean call duration

p95 latency

–

95th percentile duration

Max latency

–

Slowest call duration

Cost saved

–

Estimated, at input token price

Energy saved

–

Estimated, from tokens saved

Tokens processed

–

Total input tokens, pre-compression

Decisions stored

–

Cross-session memory (all time)

PII items masked

–

PrivacyGuard — 33 countries, GDPR/AI Act aware

Tokens saved over time

Compression savings by day, selected range

reads ~/.synthelion/savings.jsonl

By tool

Share of tokens saved per tool

updates on refresh

By content type
TypeTokens saved
Loading…
Cleanup
Delete session records older than
Sessions
one row per synthelion-mcp / CLI process
Session PID Calls Tokens saved Tools used First activity Last activity
Loading…
Recent requests
–
Time Tool Type Before After Saved Efficiency Latency Session
Loading…
Cleanup
Delete decisions older than
Recent decisions (session memory)
–
  • Loading…
This node

Multi-node cluster role — standalone, master, or a slave joined to a master

standalone

Node ID

–

Master URL

–
Join a master

Copy the master's node URL and token from its own Cluster page, then join from here

Deploy files

Docker/Kubernetes templates for this cluster topology — no token baked in, fill in ${SYNTHELION_NODE_TOKEN} at deploy time

Download docker-compose.yml Download k8s manifest
Compression

Default level used when a caller (CLI, MCP tool) doesn't specify one

Project Wiki

Default detail level (1-4) for `synthelion wiki` / the MCP tool when not specified per call

Dashboard

Host/port this web server binds to, and how the browser gets live updates

Requires restarting the dashboard process to take effect — the server is already bound to the current host/port and won't rebind itself.

Cluster storage

Session/analytics store and cross-session memory (RAG) backend

Storage status
read-only, current process

Sessions on disk

–

Decisions stored

–

Ledger records

–

Active memory backend

–
PrivacyGuard

PII detection & masking across 33 countries, prompt-injection screening — a direct port of Caveman.PrivacyGuard (C#)

Settings

Active on the compress hook by default — every toggle here is adjustable

When the toggle above is on, the UserPromptSubmit hook rejects the prompt outright (with the full PII/privacy note and AI Act notice as the reason) instead of masking-and-continuing.

Turning "Enabled" off restores exactly the pre-1.2.2 behavior (no privacy pre-pass at all).

Whitelist

Exact-match values PrivacyGuard should never flag or mask (e.g. a public support email)

  • No whitelisted values yet.
Live tester

Paste text to see PII detection and prompt-injection screening side by side — nothing here is persisted

PII / Privacy

Score: – · Risk: –

Categories: –

Compliance: –

Masked: –

Prompt injection

Score: – · Risk: –

Categories: –

Local privacy/compression proxy

Point an agent's ANTHROPIC_BASE_URL / OPENAI_BASE_URL at this proxy for real, enforced masking + compression — no MCP or hook support required on the agent's side. Off by default; runs independently of the MCP integrations, which keep working exactly the same whether this is on or off.

Status

State: –

Address: –

PID: –

Upstreams
Custom routes

Checked before the built-in Anthropic/OpenAI/Gemini routing, first match wins

LabelPath prefixUpstream
Loading…

One explicit, on-demand request when you click "Fetch list" — never automatic, same policy as the PyPI update check.

Reliability

Failover (max 10 backup upstreams)

Loading…

Circuit breaker

Rolling-history compression

Once a conversation reaches the threshold, older turns compress at "aggressive" while recent ones stay at the configured default

CCR — reversible compression

Cache originals so an agent can ask for the full detail back via synthelion retrieve --token ...

Response cache

Exact-match only (not semantic/ML) — an identical request within the TTL is served without calling the provider again

Budget & output shaping
Recent calls
Metadata only — prompts/responses are never logged
TimeMethodPathUpstream StatusDuration SavedResult
Loading…
Firewall & WAF

Request pattern inspection (SQLi, XSS, path traversal, command injection, scanner probes, bad user agents), IP allow/block lists, auto-ban and rate limiting — a direct port of the digitalsolutions WAF, protecting the dashboard's network surface

Settings

Default is detect-only — nothing gets blocked until you turn on Block mode

Rule categories

Auto-ban

Rate limit (firewall)

IP rules

Manual allow/block entries — auto-ban and rate-limit bans also show up here

IP Kind Reason Expires
No IP rules yet.
Recent events

Most recent detections/blocks — newest first

Time IP Request Category Rule Severity Action
No events yet.
Doctor
  • Not run yet.
Version

Running v—

Only checks PyPI when you click the button — Synthelion makes no network calls otherwise.

–

Dashboard administrator

Account

Change the dashboard login

Account information
  • Username:  –
  • Synthelion version:  –
  • Session store:  –
  • Memory backend:  –
Recent notifications
  • Loading…
Notifications
  • Loading…

Local dashboard — reads from ~/.synthelion/. Settings/account changes write to disk.