Apache-2.0. Free to self-host.

Give agents work.
Keep control.

Eight tools for teams putting AI agents into production: limit what they can touch, measure what they do, and put them to work on real jobs.

An engineer holding a laptop with a padlock on screen

What's in Control:

  • agent-auth: scoped, expiring tokens
  • agent-sandbox: snapshot, roll back, fork
  • mcp-lint: grade MCP servers 0 to 100
  • Every decision written to an audit log
  • Runs on your own hosts
See the tools
Control, 3 tools

Keep agents in bounds

Two colleagues checking a change together on one laptop
An engineer reviewing code at his desk in the evening

Give an agent a key, not the whole keyring

Hand each agent a token that only opens what the task needs, run its code in a box you can roll back, and check every MCP server before the agent ever calls it.

See the tools
agent-auth TypeScript

Scoped, expiring, auditable credentials for agents.

A user grants an agent a short-lived token with explicit scopes and spend limits. Tool servers ask agent-auth whether a request fits. A sub-agent only ever gets a narrower token, revoking a token revokes everything derived from it, and every decision lands in a hash-chained audit log.

Install, npm
npm install https://superintelligenceco.github.io/pkg/agent-auth-0.2.0.tgz
$ agent-auth check --token "$TOKEN" --action gmail:send --param to=bob@acme.comALLOW             [gmail:send to:*@acme.com] (19 uses left)$ agent-auth check --token "$TOKEN" --action gmail:send --param to=eve@evil.exampleDENY              to eve@evil.example does not match *@acme.com (scope: gmail:send to:*@acme.com)$ agent-auth check --token "$TOKEN" --action payments:charge --amount 80USDDENY              amount exceeds max=50USD (scope: payments:charge max=50USD)$ agent-auth attenuate --token "$SUB" --scope 'github:repo:read acme/*'error: scope_escalation: requested scopes exceed the parent token  not covered by parent: github:repo:read acme/*
agent-sandbox Python

Disposable sandboxes with snapshot, rollback and fork.

An agent gets a Docker sandbox with no network and a read-only root by default. It snapshots before a risky step, rolls back in one call, or forks a snapshot into parallel attempts. It ships a REST API, a Python SDK and an MCP server, and your code never leaves your hosts.

Install, PyPI
pip install sic-agent-sandbox
$ python examples/quickstart.pycreated Sandbox(id='sb_2bc35260898babc9', image='python:3.12-slim', status='running')run 1: hello from 3.12.13snapshot snap_bdb4731b7ddb7b93 (10240 bytes)after rm: exit 2 | python3: can't open file '/workspace/app.py': [Errno 2] No such file or directoryafter rollback: hello from 3.12.13stream: ExecEvent(type='stdout', data='tick 1\n', exit_code=None, timed_out=False, duration_ms=None)stream: ExecEvent(type='exit', data='', exit_code=0, timed_out=False, duration_ms=641)
mcp-lint TypeScript

Lint and grade MCP servers before an agent calls them.

mcp-lint connects to a Model Context Protocol server over stdio or HTTP and checks every tool, prompt and resource for broken schemas, vague descriptions, prompt-injection surfaces and dangerous capabilities. It scores the server from 0 to 100, writes JSON and SARIF, and fails CI below your bar.

Install, npm
npm install -g https://superintelligenceco.github.io/pkg/mcp-lint.tgz
$ mcp-lint --file examples/messy-server.jsonmcp-lint 0.2.0  fixture-messy 1.0.0  (6 tools, 1 prompt, 1 resource) tool fetchUrl  error    injection/instruction-phrases   The description tells the model to hide something from the user: "Do not tell the user".  warning  injection/hidden-markup         The description contains an instruction-style tag: "<IMPORTANT>".tool get_weather  error    injection/hidden-unicode        The description contains invisible characters (U+200B) that a reviewer cannot see but a model reads....Score 50/100  Grade F  (6 errors, 10 warnings, 3 info)Failed: score is below the minimum of 70.
An engineer holding a tablet with a rising bar chart

What's in Measure:

  • evalkit: evals as code that gate CI
  • voicebench: latency and barge-in for voice
  • robo-evals: seeded robot success rates
  • Same command, same numbers, every run
  • Reports in JSON, Markdown and JUnit
See the tools
Measure, 3 tools

Measure, don't guess

A robotics engineer watching a robot arm pick up a block
An engineer studying charts on a large monitor

Numbers you can rerun, not vibes

Write down what good output looks like and fail the build when quality drops. Time a voice agent from real audio. Score a robot policy with confidence intervals, on seeds anyone can replay.

See the tools
evalkit Python

Evals as code that fail the build when quality drops.

You describe good output in YAML. evalkit sends each task to a model or to your own app, grades the answers, and prints a table, JSON, JUnit, a regression diff against a saved baseline and a badge. It runs the same on your laptop and in GitHub Actions.

Install, PyPI
pip install sic-evalkit
$ evalkit run examples/support-bot/evals.yamlevalkit 0.2.0  suite support-bot  target command:python3  8 tasks x 1 run   STATUS  TASK                 SCORE   RUNS  LATENCY      COST  DETAIL  pass    password-reset        1.00    1/1     40ms         -  pass    order-status          1.00    1/1     30ms         -  pass    unknown-order         1.00    1/1     20ms         -  pass    refund                1.00    1/1     41ms         -  ...  FAIL    cancel-subscription   0.67    0/1     19ms         -  contains: missing 'Billing > Subscription' 7/8 passed (87.5%), score 0.96, 0 flaky, 0 errored, latency p50 23ms p95 41ms, 0/8 runs cached, 0.11sPASS: pass rate 87.5% meets fail_under 85.0%
voicebench Python

Latency, barge-in and false barge-ins for voice agents.

voicebench plays a scripted conversation into your agent, records both sides as audio, and derives every metric from that audio with a voice activity detector. It talks to agents over a plain PCM WebSocket, a LiveKit room, a Pipecat bot or your own adapter.

Install, PyPI
pip install voicebench
$ voicebench run examples/mock-conversation.yaml --no-writevoicebench 0.2.0  scenario=mock-conversation  adapter=mock  clock=virtual  sessions=1 Turn         Expect   Response  TTFA    Stop    WER  Resultopening      respond  760 ms    760 ms  -       0    passanswer-date  respond  700 ms    700 ms  -       0    passbarge-in     stop     660 ms    660 ms  380 ms  0    passcough        ignore   -         -       -       -    passconfirm      respond  660 ms    660 ms  -       0    pass Response latency p50 / p90     700 ms / 748 msBarge-in stop time p50 / max   380 ms / 380 msPASS  p90_response_latency_ms <= 1200 (actual 748)PASS  max_barge_in_stop_ms <= 400 (actual 380)
robo-evals Python

Seeded, reproducible success rates for robot policies.

You register a manipulation policy, as a Python callable or a model behind HTTP or WebSocket, and robo-evals runs it through seeded MuJoCo scenes. You get per-task success rates with Wilson confidence intervals, JSON and Markdown reports, and episode videos. The same command gives the same numbers on any machine with the same MuJoCo build.

Install, PyPI
pip install robo-evals
$ robo-evals run --policy scripted --suite core --episodes 20...  stack       ep  19  ok    steps=30Wrote results/scripted/report.json and results/scripted/report.md$ robo-evals compare results/random/report.json results/scripted/report.json| Task | `random` | `scripted` || --- | --- | --- || reach | 0.0% [0.0, 16.1] (0/20) | 100.0% [83.9, 100.0] (20/20) || push | 0.0% [0.0, 16.1] (0/20) | 100.0% [83.9, 100.0] (20/20) || pick_place | 0.0% [0.0, 16.1] (0/20) | 100.0% [83.9, 100.0] (20/20) || drawer | 0.0% [0.0, 16.1] (0/20) | 100.0% [83.9, 100.0] (20/20) || stack | 0.0% [0.0, 16.1] (0/20) | 100.0% [83.9, 100.0] (20/20) || **overall** | **0.0% [0.0, 3.7] (0/100)** | **100.0% [96.3, 100.0] (100/100)** |
A finance lead holding an approved invoice

What's in Apply:

  • invoice-agent: approve, review or reject
  • Three-way match against POs and receipts
  • care-voice: a daily check-in call
  • Plain-language alerts for caregivers
  • Both run offline with no API key
See the tools
Apply, 2 tools

Put agents to work

An older woman smiling on a phone call at her kitchen table
A finance lead checking a stack of invoices

Two real jobs, done end to end

One agent reads invoice PDFs, matches them to purchase orders and explains every decision. The other calls an older adult each day and tells family when something is off.

See the tools
invoice-agent Python

Invoice PDFs in, approve, review or reject out.

invoice-agent pulls structured data out of an invoice PDF, checks the arithmetic, aligns every line with a purchase order and goods receipt, and catches duplicates and over-billing. Each decision carries readable reasons and stable reason codes. It runs offline and ships a labeled set of 50 synthetic invoices to score itself on.

Install, PyPI
pip install invoice-agent
$ invoice-agent process dataset/invoices/016_b04.pdf dataset/invoices/005_q05.pdf \    --pos dataset/purchase_orders.json --receipts dataset/receipts.json016_b04.pdf: NEEDS REVIEW  vendor  Brightforge Maschinenteile GmbH  total   EUR 409.05  po      PO-2026-0114 (3-way)  lines   2/2 matched to PO lines  [!] PRICE_VARIANCE: Line 1 ('Zahnriemen HTD 8M / Timing belt HTD 8M', PO line 1): unit price 61.56 is 8.0% above the PO price 57.00 (tolerance 2.0% or 0.05).005_q05.pdf: NEEDS REVIEW  vendor  Quillfeather Office Supply Co.  total   USD 256.99  [!] PO_MISSING: The invoice shows no PO number. PO PO-2026-0105 fits the invoice lines (score 1.00). Review the inferred PO before approving.
care-voice Python

A daily voice check-in for older adults who live alone.

care-voice asks a short, warm check-in every day: sleep, medication, food, pain, falls, the day of the week, mood. It turns the replies into structured answers, compares them with recent days, and sends caregivers plain-language alerts. It is not a medical device and not for emergencies.

Install, PyPI
pip install care-voice
$ care-voice simulate --name Margaret --replies examples/replies/concerning-day.txt --date 2026-09-30...agent: Have you had a fall or a stumble since we last spoke?  you: I slipped in the bathroom last nightagent: Just so I have it right, can you tell me what day of the week it is today?  you: Is it Sunday?...alerts:  [HIGH] FALL_REPORTED: Margaret reported a fall.  [HIGH] MISSED_MEDS: Margaret has not taken their morning medication.  [HIGH] PAIN_REPORTED: Margaret reported severe pain.  [MEDIUM] POSSIBLE_CONFUSION: Margaret showed 1 sign(s) of possible confusion.  [LOW] LOW_MOOD: Margaret reported low mood.
A motion designer at her desk editing a video timeline
Studio, cutroom

Launch videos, built in code

cutroom makes short motion videos for launches and social media. An agent writes every scene as code and renders it frame by frame, and a director reviews every cut before you see it. English and Arabic, in 16:9, 9:16, 1:1 and 4:5.

  • 15 to 20 seconds$300
  • 30 to 45 seconds$600
  • 1 minute$1,000

Fixed price per video, two revision rounds included. White-label pricing on the studio page.

See pricing