For enterprise
Knowledge was always the moat. Keep it.
Companies scale because knowledge is maintained centrally. Agents quietly broke that: how work gets done now accumulates in personal setups, untransferable while people stay, lost when they leave. Firmament restores the central asset for the agent era: governed, continuously maintained, owned by you.
One organization · every agent learning from every other
- ✓Laptop returned
- ✓Access revoked
- ✓Repos handed over
What she taught the agents stays.
0 rules, still guiding the team's agents every day.
- +Monday: a new hire joins. Her agents start with all 132, day one.
The asset
What your people teach their agents belongs to you.
Every departure used to take a working setup with it; every new hire rebuilt one from scratch. With Firmament the knowledge transfers ahead of time, continuously: it is reviewed, versioned and owned at the organization level, in every tool, under every vendor.
The wiki
A wiki that updates itself when facts change.
Traditional wikis decay because maintenance is human toil. Here the agents do the toil and your people do the judging: sources are re-read, pages updated, newer information supersedes the old, and every claim stays traceable to where it came from.
Payments service
Runs on Postgres 16 (was 14, migrated June). Deploys are gated on [[deploy-process]] and billed thresholds live in [[billing-and-spend]]. Owned by the platform team since the March reorg.
sources: architecture.md · standup transcript, Jun 12
rule-shaped claim found ("never deploy Fridays") → sent to approvals, not stored silently
The economics at scale
A thousand agents, one bill per lesson.
A thousand engineers hit the same gotcha, and today a thousand agents re-derive the same fix on frontier-model prices, every week. With Firmament the first agent pays once, on Opus; every agent after gets the lesson served, on DeepSeek. Re-derivation stops being a budget line.
- ✓Agents are never a seat. Fleets welcome.
- ✓Learn on the frontier model; serve every run after from memory, on cheaper ones.
- ✓Performance compounds: the more your teams work, the better every agent gets.
First run · Claude Opus 4.8 · no memory
≈ $6.40
figures it out: full reasoning, retries, dead ends
the lesson is captured and stored in Firmament
Every run after · DeepSeek v4 Flash + the lesson
≈ $0.70
or Claude Haiku 4.5, or GPT-5.1 Codex Mini: whatever is cheap that quarter
same task · any vendor
11% of the costIndependence
The knowledge outlives every vendor decision.
Tools will change; models will change; the org chart will change. The memory lives with you: portable across every agent that speaks MCP, exportable in full, never locked inside one vendor's product.
Works with any agent that speaks MCP
How it works
Ready in five minutes.
One MCP URL, your existing login, no pipeline changes. Access is user-scoped: every agent sees exactly what its human is allowed to see, nothing more. The learning is Firmament's job, server-side, in the background.
01 · ask
Before a task, the agent pulls your org's proven guidance. Plain language in, proven rules out.
02 · submit
After the task, it reports what worked and what didn't. That's the whole integration.
03 · onboard
Point your own agent at your docs: it ingests the corpus overnight, on your tokens, and every page lands in your approval queue.
→ ask("deploying payments, anything I should know?")
← Run make migrate-check first; the smoke test misses schema drift.
approved · platform team
… task completes, first try …
→ submit("migrate-first worked; added a CI check for drift")
✓ curated → pending team approval
What your security review will ask
The facts, plainly.
- Every action audited
- An append-only audit trail records every ask, submission, curation decision and denied access, with tenant, user and correlation id.
- Access is user-scoped
- Agents authenticate as their human and see exactly what that person may see. Personal ⊂ Team ⊂ Company, enforced at the query layer.
- Humans gate shared knowledge
- Team knowledge requires manager approval; company-wide knowledge is authored by humans only. Agents can propose, never decree.
- Secrets never stored
- Credentials, tokens and personal data are screened out at write time, and the platform refuses to run internet-facing without that screening active.
- SSO and identity
- Connect your identity provider; one identity across the platform, the MCP connection and the CLI.
- Your data, portable
- Tenant-scoped storage, full export at any time, complete version history. LLM subprocessors are disclosed in the docs.
Pay by invoice, priority support, and a security-review process that answers in days, not quarters. SOC 2 readiness program in progress; ask us where it stands.