Key facts
| Agent runtime | Function calling loop on /v1/chat/completions (live) |
| Tools | Matter search, clause extraction, precedent lookup and draft creation |
| Models | 30+ models, route extraction and long-document synthesis separately |
| Long context | plugsky-longctx for comparing agreements and building term summaries |
| Access control | Scoped keys with matter-level permissions and SSO/SCIM |
| Audit | Per-turn tool calls and inference metadata exportable to SIEM |
| Deployment | Region-locked planes, VPC, on-prem and air-gapped options |
| Roadmap endpoints | Assistants-style managed endpoints are coming soon; the tool loop is live now |
TL;DR
- Scope the agent to research, extraction and drafting — lawyers own advice and sign-off.
- Enforce matter walls inside tools, not in prompts, so conflicts stay protected.
- Extract clauses with structured output so downstream review is comparable.
- Keep an audit trail of every retrieval and draft for privilege and risk reviews.
- Evaluate on closed matters where outcomes are known before widening access.
How it works, step by step
- Pick one bounded task — clause extraction, first-pass due diligence or precedent search — and define its allowed actions.
- Expose matter files, clause libraries and precedent systems as tools with typed schemas and server-side matter-wall checks.
- Implement the loop: send the instruction with a tools array, execute each tool_call, append results, and stop when a draft is ready.
- Require structured output for extractions — clause type, position, risk flag, source paragraph — so reviewers compare like with like.
- Add lawyer review gates before anything leaves the team, and label all output as a draft for verification.
- Log tool calls, retrieved documents and model metadata per turn, with retention aligned to matter policy.
- Evaluate on closed matters, measuring extraction recall and how often reviewers accept or edit drafts.
Try it yourself
Open the AI data residency checklist →
Tools for legal work
Legal tasks decompose well into tools: search matter files, pull a clause from the library, find precedent, compare two versions, and create a draft. Each tool returns citable, structured results, and the agent's job is to plan the sequence and write the summary. That keeps the model in the role it performs best — reading and reorganising text — while deterministic systems handle access and storage.
Plugsky's live function calling supports the loop directly: a tools array on /v1/chat/completions, tool_calls in the response, results appended, iterate until done. OpenAI compatibility means the integration reuses your existing SDK patterns and stays inside your document management boundary.
Privilege, confidentiality and review
Confidentiality is the design constraint, not a feature to add later. Matter walls must be enforced in tool code against the signed-in lawyer's matters, and prompts should carry only the passages needed for the task. Anything leaving the environment must be a deliberate, logged decision.
- Matter scoping: resolve the user's matters server-side; never accept a matter ID supplied by the model.
- Minimisation: redact third-party personal data where the task does not need it.
- Privilege: keep the agent's outputs labelled as drafts and store them where privilege rules require.
- Residency: region-locked planes for cloud, or VPC, on-prem and air-gapped when files cannot leave the firm.
Evaluation and rollout in a legal team
Legal quality is measured against closed matters, not benchmarks. Build a set of documents where the correct clause positions, dates and obligations are known, then score extraction recall, citation accuracy and how often reviewers accept a draft without material edits. Track false negatives on risk flags, because a missed indemnity matters more than a stylistic miss.
Start with one practice group and one document type. Use smaller models for extraction and classification, and stronger models for cross-document reasoning — 30+ models behind one endpoint makes that a configuration decision. Expand only when the review data shows the assistant reduces time without reducing scrutiny, and keep the human sign-off step explicit in every workflow.
Honest comparison
| Concern | Plugsky legal agent | Manual review | Legal AI point tool |
|---|---|---|---|
| Scope | Research, extraction and drafting | All work | Usually one document type |
| Access control | Matter walls enforced in tool code | Physical and system access | Vendor-defined |
| Audit | Per-turn tool and model logs | Time records and notes | Vendor-controlled logs |
| Model choice | 30+ models behind one endpoint | Human judgement | Fixed model |
| Deployment | Region-locked planes, VPC, on-prem, air-gapped | On premises | Usually cloud-only |
Frequently asked questions
Can the agent give legal advice?
No. It researches, extracts and drafts; qualified lawyers review and own the advice. Build the workflow so every output is labelled as a draft requiring verification.
How do matter walls work?
Your tools resolve the signed-in lawyer's matters server-side and filter all retrieval accordingly. The model never receives a matter identifier it can substitute for another.
Is function calling live?
Yes. OpenAI-style function calling and streaming are live on chat completions, which is sufficient to build the tool loop today.
What can it do well today?
Clause extraction, document comparison, first-pass due diligence checklists, precedent search and draft summaries — all tasks where a human verifies before use.
How do we protect privileged material?
Keep documents in your systems, send only needed passages, scope keys per team, and pin deployment to a region or run on-prem where privilege rules require it.
How should we evaluate it?
Use closed matters with known answers, score extraction recall, citation accuracy and reviewer edit rate, and treat missed risk flags as the most important failure.
Where do we start?
One practice group, one document type and one workflow. Measure time saved and review quality before expanding to the next use case.