Agent context
File reads, search results, tool output, and thinking accumulate with every turn.
Parsec · Context for coding agents
Less context.
More capability.
Give your agent the context it needs to move forward. Parsec trims unused tool output and thinking, so your agent spends less, runs faster, and has more room to work.
Works withClaude Code · Codex · opencode
Continue with Google or GitHub. Set up from your dashboard.
Learned curator
Parsec scores each chunk and removes context the agent is unlikely to use.
Original content
Kept content stays byte-exact. Your prompts and the agent’s answers stay untouched.
Measured results
Less context to process. Lower cost, faster runs, and more solved tasks in the benchmark.
SWE-bench Verified
Cheaper. Faster. More solved.
- Lower cost
- −39%
- $89.65 vs $147.30 total cost
- $1.45 per solved task. The lowest of all six arms tested.
- Less time
- −25%
- 10.8 hours vs 14.4 hours
- The same task set, completed 3.6 hours sooner.
- More tasks solved
- 62/100
- Up from 57 without compression
- Five more working fixes, verified by the official grader.
Input tokens
−54%
Without compression312.2M
With Parsec144.8M
Output tokens
−43%
Without compression3.00M
With Parsec1.71M
Dasein's benchmark: 100 SWE-bench Verified tasks, one fixed model (Claude Sonnet 4.6), the official Docker grader, and cache-aware pricing. Only the compression layer changes. Results are specific to this run.
Less noise, same source
Most context never gets used.
Across 28,487 real agent traces, 78.4% of trim-eligible context was never used. Parsec's curator scores chunks to predict what your agent will need next.
78.4%
of eligible context went unused in the trace study
Unused context by type
- Search results
- 98.2%
- Context & plans
- 84.9%
- File reads
- 82.2%
- Thinking
- 69.7%
- Command output
- 68.1%
- Web content
- 67.9%
Your prompts & the agent's answers
Never touched
Keep the original words
Retained content stays byte-exact. No rewording or summaries to inspect. Your prompts and the agent’s answers are never altered.
Curate the whole turn
Trim tool output and thinking, then prune the tool roster each turn. The curator learns what matters from real agent traces.
Use your own credentials
The proxy runs locally. Model requests go directly to your provider with your existing credentials. Hosted scoring handles context selection.
Read the receipts
A local ledger records each request. Savings use measured token counts where available; Codex reports actual billed usage without estimating savings.
How it works
Your agent. A lighter context.
Keep your tools and your workflow. Parsec sits between your agent and its model, curating context before each request.
01
Connect your agent
Sign in, then follow the setup in your dashboard. One local proxy works with Claude Code, Codex, and opencode using your existing model credentials.
02
Keep what matters
A curator learned from real agent traces scores tool output and thinking before each turn. It drops unused context and prunes the tool roster.
03
See the difference
Inspect usage in a local ledger. Supported providers show measured savings per request; Codex records actual billed usage without an estimated savings baseline.
Pricing
Start with your next session.
Pro
Start here
Plugin + local proxy
Free
- Full learned curator via the hosted scoring API
- Tool-schema pruning
- Governor
Team
In our cloud
Hosted BYOK gateway
Coming soon
- Your API keys
- Zero local install
- Centralized reporting
Enterprise
In your VPC
Deployed in your cloud
Talk to us
- The full stack, including the scoring API
- Licensed, containerized checkpoints
- Fine-tuning on traces inside your network
Which coding agents does it work with?
Claude Code, Codex, and opencode. One local proxy serves all three; the dashboard provides setup instructions for your agent.
Does my code go to your servers?
Hosted scoring receives context chunks over TLS. Model requests go directly to your provider. Enterprise can run scoring inside your VPC.
Can I use my Claude or ChatGPT subscription?
Yes. Local connections use your existing sign-in. The hosted Team gateway accepts API keys only.
What happens if the scoring API is down?
Requests pass through uncompressed, so you can keep working. These events are recorded in the ledger.
How are savings calculated?
Original token counts are compared with billed usage per request. Codex reports billed usage only, because no equivalent token-counting endpoint is available.
Does Parsec train on my data?
Not by default. Telemetry is opt-in, starts disabled, and can be previewed before upload.