Agents

An agent is a role filler that happens to be software.

It belongs to a team. It receives a named role, a clear focus, current organizational context, and explicit limits. It does not enter through a workspace-wide credential and infer its own authority.

A charter editor creates the agent from the team's Relay tab, assigns it to one or more roles, and gives the agent access to a compiled operating charter.

The model shapes the agent. Agent access is read-only. The agent cannot rewrite the organization that authorizes it.

Connect an agent

Open the team charter.
Choose Relay → Create agent.
Assign the agent to a role.
Copy the token. It is shown once.
On the computer running the agent, in the folder that agent works from, run:

npx particulars

Paste the token when it asks. The command checks it with Particulars — a mistyped or revoked token fails right there rather than inside the agent later — and then leaves two things behind:

A skill file, skills/Particulars/SKILL.md. Any AI agent that reads skills from that folder picks it up on its own and uses it to load the roles and charters it is working under, refreshing CHARTER.md every time it starts.
The token, saved under that agent's own name, so you never hand it over again.

It then offers to add the one line that makes the agent look:

Read CHARTER.md before acting - it is your operating charter from Particulars.

If the folder already has an AGENTS.md or a CLAUDE.md, it asks before appending to it. If it has neither, it offers to write an AGENTS.md. Decline and it prints the line for you to paste wherever your agent reads its standing instructions.

Several agents on one computer

Each agent keeps its own token, named after it: an agent called Scout is saved as PARTICULARS_AGENT_TOKEN_SCOUT_3F2A, where the last part comes from that agent's id so two agents sharing a name never share a token. Connect each one from the folder it works in — pass --dir if you are setting them up from elsewhere — and each reads only the roles it fills. Decline the save and the command prints the line to set that variable yourself instead.

For unattended or scripted setup, npx particulars --help lists the flags: --token, --dir, --save-token, --no-save-token, --write-instructions, --no-write-instructions.

Connecting an agent that isn't an openclaw skill

The connector scaffolds one specific shape of agent. If your agent runs a custom harness, a CI job, or anything else that reads context over HTTP instead of loading a skill file, call the endpoint directly:

curl -s \
  -H "Authorization: Bearer $PARTICULARS_AGENT_TOKEN" \
  https://particulars.team/api/agent-context

The endpoint returns Markdown suitable for operating context. ?format=json returns a structured payload.

Use If-None-Match with the most recent ETag. A 304 response means the cached context is still current.

What the agent reads

For each role it fills, the agent receives:

the team's purpose
the team's focus areas
relevant measures
the role's focus and details, written in second person
the team's decision rights as behavioral rules
required consultation and information relationships
a short roster of people and agents on the team
the team's guidelines
standing confirmed decisions

The context tells the agent what it is for, what it may do, when it must ask, where its output goes, and who remains responsible.

What the agent cannot do

The agent context endpoint is read-only.

An agent cannot:

edit the charter
change its own role
expand its own authority
view organizations or teams outside its scope
retain access after revocation

Revoking an agent invalidates its token and removes it from its assigned roles.

Pricing

Agents are included on paid organization plans. They are not billed as human seats, and ordinary usage does not create per-action overage charges.

Current plan names and limits are listed on Pricing.