Point a team of AI agents at a question that never stops mattering — what competitors are shipping, how the market talks about you, what the category is searching for — and they will watch it continuously, agree on what they found, and keep a permanent record you can prove wasn't edited.
You describe the goal. The fleet builds itself.
Agents forget. Ask one to track something over months and you get a fresh, confident, slightly different answer every session — with no way to tell what changed because the world moved and what changed because the model did.
This fixes that with three moving parts:
Many lenses, one thread. Several agents watch the same topic from deliberately different angles. Their disagreement is the signal — one lens sees a pricing page change, another sees the press narrative that contradicts it.
A shared memory, kept small. Findings go into one cross-agent index every agent reads and writes. It stays curated on purpose: notes get merged, contradictions get resolved, and the merged result supersedes the noise.
A record that can't quietly change. Each merged artifact is stored on Filecoin, where its address is the hash of its contents. Retrieving it re-verifies it. If the index and the archive ever disagree, the archive wins — and you can check that yourself with a plain HTTP request, no account required.
The result is institutional memory that compounds instead of resetting.
You: "Set up a pipeline."
Agent: What should the fleet watch, and what decision does it serve?
→ e.g. know within a day when a competitor changes pricing
Which angles? Two lenses disagree usefully; one just repeats itself.
→ pricing pages + changelogs · launch announcements + press
How fresh does it need to be?
→ daily (hourly costs ~24× the model usage to learn the same thing)
What counts as load-bearing?
→ cites a primary source; commentary alone gets dropped
Agent: Here's the spec. One supervised run first, then I'll schedule it.
No config file to learn. The fleet-creator skill interviews you, offers worked examples to react to, validates the result, proves it works once, then puts it on a schedule — and tears it back down cleanly when you're done.
| A standing answer | One current, merged artifact per topic — not a pile of session notes |
| Provenance | Every claim cites a source; every archived version names the version it replaced |
| Proof | Content-addressed storage with continuous on-chain proofs; verify by downloading |
| Portability | The knowledge outlives the agent, the vendor, and the session |
| A predictable bill | Storage is free on testnet. Model usage is the real cost, and cadence is yours to set |
Works on Linux, macOS, and Windows.
1 · Install the skills — no clone required:
npx skills add FIL-Builders/filecoin-clawdi-fleet -g2 · Wire the fleet — tell your agent "set up the fleet". It detects what you have installed and wires every layer, verifying as it goes. One human step, done once:
clawdi vault set FILECOIN_PRIVATE_KEY --prompt # a dedicated low-value wallet
foc-cli wallet init --keyRef clawdi:FILECOIN_PRIVATE_KEY # stores a reference, never the key
foc-cli wallet balance --json # keySource: "keyRef" + address3 · Create a pipeline — tell your agent "set up a pipeline" and answer its questions.
Full walkthrough: docs/setup.md → docs/pipelines.md.
Storage is paid for by a wallet the agents can use but never see. The key lives in the
Clawdi vault; foc-cli stores only a reference to it and resolves the value into memory
for the single command that needs it.
The honest boundary: the key never reaches an agent's context, durable state, or argv. That is credential hygiene, not OS-level isolation — anything running as the same user can resolve the same reference. Use a dedicated low-value wallet.
Everything defaults to the free Calibration testnet (chain 314159). Mainnet spends real funds and requires explicit human confirmation.
| Skill | Role |
|---|---|
| fleet-creator | Interviews you, designs the pipeline, instantiates it, tears it down |
| fleet-setup | Detects agent × location × layer and wires each one |
| memory-researcher | The compounding research loop — structured notes that build on each other |
| memory-consolidator | Merges notes into one verified artifact, archives it, prunes the rest |
Installing lands real skill directories in ~/.agents/skills/ and wires the agents
skills.sh recognises (Claude Code, Cursor, Amp). OpenClaw, Hermes and Codex import from
there by their own route — per-agent commands, hosted boxes included, are in
fleet-setup §3.
Storage commands come from the upstream foc-cli skill.
| Path | What |
|---|---|
| docs/setup.md | Fleet setup from a fresh Clawdi account |
| docs/pipelines.md | What a pipeline is, how to run and watch one |
| docs/memory-demo.md | The memory and proof mechanics, live in ten minutes |
| docs/troubleshooting.md | Observed symptoms and fixes |
| AGENTS.md | Operating rules — auto-loaded by agents working in this repo |
| prompts/ | Researcher and archiver prompt templates |
| skills/ | The skills — the installable product |
| scripts/ | check-skills.sh, the consistency gate for skills/ |
Apache-2.0 OR MIT, matching upstream foc-cli.